2025-12-04T08:52:49.4241939Z Current runner version: '2.329.0'
2025-12-04T08:52:49.4244917Z Runner name: 'linux.rocm.gpu.gfx942.1.b-gwk9b-runner-xf6tf'
2025-12-04T08:52:49.4245349Z Runner group name: 'default'
2025-12-04T08:52:49.4245772Z Machine name: 'linux'
2025-12-04T08:52:49.4246922Z ##[group]GITHUB_TOKEN Permissions
2025-12-04T08:52:49.4247994Z Contents: read
2025-12-04T08:52:49.4248235Z Metadata: read
2025-12-04T08:52:49.4248472Z ##[endgroup]
2025-12-04T08:52:49.4249449Z Secret source: Actions
2025-12-04T08:52:49.4249780Z Prepare workflow directory
2025-12-04T08:52:49.4485314Z Prepare all required actions
2025-12-04T08:52:49.4504641Z Getting action download info
2025-12-04T08:52:49.9309512Z Download action repository 'pytorch/pytorch@main' (SHA:ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32)
2025-12-04T08:52:54.3223372Z Download action repository 'pytorch/test-infra@main' (SHA:39aa74d619174326f4e2fb0e216151c2f29d9ffd)
2025-12-04T08:52:55.6033336Z Download action repository 'actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02' (SHA:ea165f8d65b6e75b540449e92b4886f43607fa02)
2025-12-04T08:52:56.7676746Z Download action repository 'aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722' (SHA:ececac1a45f3b08a01d2dd070d28d111c5fe6722)
2025-12-04T08:52:57.8287768Z Getting action download info
2025-12-04T08:52:58.0194473Z Download action repository 'actions/checkout@v4' (SHA:34e114876b0b11c390a56381ad16ebd13914f8d5)
2025-12-04T08:52:58.9749132Z Getting action download info
2025-12-04T08:52:59.1708452Z Download action repository 'nick-fields/retry@v3.0.0' (SHA:7152eba30c6575329ac0576536151aca5a72780e)
2025-12-04T08:53:00.1321184Z Getting action download info
2025-12-04T08:53:00.3376113Z Uses: pytorch/pytorch/.github/workflows/_rocm-test.yml@refs/heads/main (ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32)
2025-12-04T08:53:00.3377992Z ##[group] Inputs
2025-12-04T08:53:00.3378142Z   build-environment: linux-noble-rocm-py3.12-mi300
2025-12-04T08:53:00.3379860Z   test-matrix: {"include": [{"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}]}
2025-12-04T08:53:00.3381890Z   docker-image: 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:00.3382167Z   sync-tag: 
2025-12-04T08:53:00.3382589Z   timeout-minutes: 300
2025-12-04T08:53:00.3382699Z   tests-to-include: 
2025-12-04T08:53:00.3382798Z   dashboard-tag: 
2025-12-04T08:53:00.3383016Z   disable-monitor: true
2025-12-04T08:53:00.3383440Z   monitor-log-interval: 5
2025-12-04T08:53:00.3383658Z   monitor-data-collect-interval: 1
2025-12-04T08:53:00.3383781Z ##[endgroup]
2025-12-04T08:53:00.3383990Z Complete job name: linux-noble-rocm-py3.12-mi300 / test (default, 3, 6, linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests)
2025-12-04T08:53:00.3612851Z ##[group]Run pytorch/pytorch/.github/actions/checkout-pytorch@main
2025-12-04T08:53:00.3613128Z with:
2025-12-04T08:53:00.3613224Z   no-sudo: true
2025-12-04T08:53:00.3613325Z   submodules: recursive
2025-12-04T08:53:00.3613423Z   fetch-depth: 0
2025-12-04T08:53:00.3613560Z env:
2025-12-04T08:53:00.3613646Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:00.3613756Z ##[endgroup]
2025-12-04T08:53:00.3671941Z ##[group]Run echo "IN_CONTAINER_RUNNER=$(if [ -f /.inarc ] || [ -f /.incontainer ]; then echo true ; else echo false; fi)" >> "$GITHUB_OUTPUT"
2025-12-04T08:53:00.3672328Z echo "IN_CONTAINER_RUNNER=$(if [ -f /.inarc ] || [ -f /.incontainer ]; then echo true ; else echo false; fi)" >> "$GITHUB_OUTPUT"
2025-12-04T08:53:00.3678629Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:00.3678779Z env:
2025-12-04T08:53:00.3678866Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:00.3678966Z ##[endgroup]
2025-12-04T08:53:00.3836926Z ##[group]Run actions/checkout@v4
2025-12-04T08:53:00.3837108Z with:
2025-12-04T08:53:00.3837229Z   ref: ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T08:53:00.3837362Z   fetch-depth: 0
2025-12-04T08:53:00.3837457Z   submodules: recursive
2025-12-04T08:53:00.3837685Z   show-progress: false
2025-12-04T08:53:00.3837821Z   repository: pytorch/pytorch
2025-12-04T08:53:00.3838002Z   token: ***
2025-12-04T08:53:00.3838094Z   ssh-strict: true
2025-12-04T08:53:00.3838193Z   ssh-user: git
2025-12-04T08:53:00.3838300Z   persist-credentials: true
2025-12-04T08:53:00.3838408Z   clean: true
2025-12-04T08:53:00.3838511Z   sparse-checkout-cone-mode: true
2025-12-04T08:53:00.3838628Z   fetch-tags: false
2025-12-04T08:53:00.3838726Z   lfs: false
2025-12-04T08:53:00.3838823Z   set-safe-directory: true
2025-12-04T08:53:00.3838942Z env:
2025-12-04T08:53:00.3839037Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:00.3839142Z ##[endgroup]
2025-12-04T08:53:00.4368902Z Syncing repository: pytorch/pytorch
2025-12-04T08:53:00.4369475Z ##[group]Getting Git version info
2025-12-04T08:53:00.4369651Z Working directory is '/home/runner/_work/pytorch/pytorch'
2025-12-04T08:53:00.4369909Z [command]/usr/bin/git version
2025-12-04T08:53:00.4370024Z git version 2.52.0
2025-12-04T08:53:00.4393005Z ##[endgroup]
2025-12-04T08:53:00.4402137Z Copying '/home/runner/.gitconfig' to '/home/runner/_work/_temp/cf1c8cd7-b46c-40d3-83d1-addcf96d6eb1/.gitconfig'
2025-12-04T08:53:00.4403459Z Temporarily overriding HOME='/home/runner/_work/_temp/cf1c8cd7-b46c-40d3-83d1-addcf96d6eb1' before making global git config changes
2025-12-04T08:53:00.4403964Z Adding repository directory to the temporary git global config as a safe directory
2025-12-04T08:53:00.4406710Z [command]/usr/bin/git config --global --add safe.directory /home/runner/_work/pytorch/pytorch
2025-12-04T08:53:00.4442217Z [command]/usr/bin/git config --local --get remote.origin.url
2025-12-04T08:53:00.4465335Z https://github.com/pytorch/pytorch
2025-12-04T08:53:00.4479954Z ##[group]Removing previously created refs, to avoid conflicts
2025-12-04T08:53:00.4483656Z [command]/usr/bin/git rev-parse --symbolic-full-name --verify --quiet HEAD
2025-12-04T08:53:00.4506914Z refs/heads/main
2025-12-04T08:53:00.4519907Z [command]/usr/bin/git checkout --detach
2025-12-04T08:53:02.0057336Z HEAD is now at ffd9b0fb4355 Resolve collective autotuning test failure on arm (#168919)
2025-12-04T08:53:02.0112430Z [command]/usr/bin/git branch --delete --force main
2025-12-04T08:53:02.0241063Z Deleted branch main (was ffd9b0fb4355).
2025-12-04T08:53:02.0247529Z ##[endgroup]
2025-12-04T08:53:02.0254290Z [command]/usr/bin/git submodule status
2025-12-04T08:53:02.0493320Z  7e1e1fe3858c63c251c637ae41a20de425dde96f android/libs/fbjni (v0.1.0-12-g7e1e1fe)
2025-12-04T08:53:02.0547344Z  4dfe081cf6bcd15db339cf2680b9281b8451eeb3 third_party/FP16 (4dfe081)
2025-12-04T08:53:02.0611291Z  b408327ac2a15ec3e43352421954f5b1967701d1 third_party/FXdiv (b408327)
2025-12-04T08:53:02.0671065Z  c07e3a0400713d546e0dea2d5466dd22ea389c73 third_party/NNPACK (c07e3a0)
2025-12-04T08:53:02.0719735Z  3ebbc93ded7285963bff932c678fa367eb393ba6 third_party/NVTX (v3.1.0-313-g3ebbc93)
2025-12-04T08:53:02.0785385Z  1d8f600fd424278486eade7ed3e877c99f0846b1 third_party/VulkanMemoryAllocator (v2.1.0-982-g1d8f600)
2025-12-04T08:53:02.1081194Z  51a0103656eff6fc9bfd39a4597923c4b542c883 third_party/XNNPACK (remotes/origin/ds/ndk-1243-g51a0103656)
2025-12-04T08:53:02.1108789Z  01aae101b9e5e94d6c16a9514c9fb8df99c93150 third_party/aiter (v0.1.1-92-g01aae101)
2025-12-04T08:53:02.1132703Z  299e5928955cc62af9968370293b916f5130916f third_party/benchmark (v1.9.3)
2025-12-04T08:53:02.1194406Z  7fe50dc3da2069d6645d9deb8c017a876472a977 third_party/composable_kernel (rocm-6.4.3-459-g7fe50dc3d)
2025-12-04T08:53:02.1283256Z  89c932f313c6437c38f2982869beacc89c2f2246 third_party/cpp-httplib (v0.26.0)
2025-12-04T08:53:02.1364529Z  f858c30bcb16f8effd5ff46996f0514539e17abc third_party/cpuinfo (f858c30)
2025-12-04T08:53:02.1400818Z  0b1577c8c83401237d601d0d0db5210506705396 third_party/cudnn_frontend (v0.5-61-g0b1577c)
2025-12-04T08:53:02.1463901Z  f88806b1e31dfa579842638740216dd41fc6c588 third_party/cutlass (v4.3.1)
2025-12-04T08:53:02.1487935Z  c0b988d39a9e47c794d699f29930ed4d7c7e13a4 third_party/fbgemm (v1.4.0-rc1-2-gc0b988d39)
2025-12-04T08:53:02.1554952Z  979702c87a8713a8e0a5e9fee122b90d2ef13be5 third_party/flash-attention (v2.7.4)
2025-12-04T08:53:02.1577186Z  a2cd1ea3b6d3fee220106b5fed3f7ce8da9eb757 third_party/flatbuffers (v24.12.23)
2025-12-04T08:53:02.1805006Z  407c905e45ad75fc29bf0f9bb7c5c2fd3475976f third_party/fmt (12.1.0)
2025-12-04T08:53:02.1877576Z  3fb5c176c17c765a3492cd2f0321b0dab712f350 third_party/gemmlowp/gemmlowp (remotes/origin/revert-87-master-135-g3fb5c17)
2025-12-04T08:53:02.1955842Z  54cbae0d3a67fa890b4c3d9ee162b7860315e341 third_party/gloo (remotes/origin/gh/c-p-i-o/1/base-37-g54cbae0)
2025-12-04T08:53:02.2102099Z  52eb8108c5bdec04579160ae17225d66034bd723 third_party/googletest (release-1.8.0-3544-g52eb8108)
2025-12-04T08:53:02.2150199Z  719d8e6cd7f7a0e01b155657526d693acf97c2b3 third_party/ideep (pytorch-rls-v3.7.1)
2025-12-04T08:53:02.2197371Z  dec1d23ca65ab069d225dfe40dea14f455170959 third_party/ittapi (v3.25.5)
2025-12-04T08:53:02.2320539Z  31f85df8fbd89c188f14ef10f1ec65379786b943 third_party/kineto (heads/main)
2025-12-04T08:53:02.2345021Z  d7770c89632329a9914ef1a90289917597639cbe third_party/kleidiai (v1.15.0)
2025-12-04T08:53:02.2366913Z  fbd8b99c2b828428947d70fdc046bb55609be93e third_party/mimalloc (v2.2.4)
2025-12-04T08:53:02.2381474Z  55f93686c01528224f448c19128836e7df245f72 third_party/nlohmann (v3.12.0)
2025-12-04T08:53:02.2585171Z  e709452ef2bbc1d113faf678c24e6d3467696e83 third_party/onnx (v1.18.0)
2025-12-04T08:53:02.2600803Z  a799f4aed9c94b765dcdaabaeab7d5e7e2310878 third_party/opentelemetry-cpp (v1.14.2)
2025-12-04T08:53:02.2616812Z  0fa0ef591e38c2758e3184c6c23e497b9f732ffa third_party/pocketfft (release_for_eigen-40-g0fa0ef5)
2025-12-04T08:53:02.2829966Z  d1eca4e4b421cd2997495c4b4e65cea6be4e9b8a third_party/protobuf (v3.7.0-rc.2-1279-gd1eca4e4b)
2025-12-04T08:53:02.2874998Z  072586a71b55b7f8c584153d223e95687148a900 third_party/psimd (heads/master)
2025-12-04T08:53:02.2919207Z  4fe0e1e183925bf8cfa6aae24237e724a96479b8 third_party/pthreadpool (0.1-144-g4fe0e1e)
2025-12-04T08:53:02.2945424Z  f5fbe867d2d26e4a0a9177a51f6e568868ad3dc8 third_party/pybind11 (v3.0.1)
2025-12-04T08:53:02.2990275Z  f45429b087dd7d5bc78bb40dc7cf06425c252d67 third_party/python-peachpy (remotes/origin/pre-generated)
2025-12-04T08:53:02.3035982Z  5a1d179df9cf652951b59010a2d2075372d67f68 third_party/sleef (3.8)
2025-12-04T08:53:02.3075034Z  2b4cd91092d335a697416b2a3cb398283246849d third_party/tensorpipe (heads/main)
2025-12-04T08:53:02.3085080Z ##[group]Cleaning the repository
2025-12-04T08:53:02.3088032Z [command]/usr/bin/git clean -ffdx
2025-12-04T08:53:02.3200274Z [command]/usr/bin/git reset --hard HEAD
2025-12-04T08:53:03.9339836Z HEAD is now at ffd9b0fb4355 Resolve collective autotuning test failure on arm (#168919)
2025-12-04T08:53:03.9406254Z ##[endgroup]
2025-12-04T08:53:03.9409034Z ##[group]Disabling automatic garbage collection
2025-12-04T08:53:03.9422080Z [command]/usr/bin/git config --local gc.auto 0
2025-12-04T08:53:03.9452248Z ##[endgroup]
2025-12-04T08:53:03.9452564Z ##[group]Setting up auth
2025-12-04T08:53:03.9454734Z [command]/usr/bin/git config --local --name-only --get-regexp core\.sshCommand
2025-12-04T08:53:03.9480665Z [command]/usr/bin/git submodule foreach --recursive sh -c "git config --local --name-only --get-regexp 'core\.sshCommand' && git config --local --unset-all 'core.sshCommand' || :"
2025-12-04T08:53:03.9694409Z Entering 'android/libs/fbjni'
2025-12-04T08:53:03.9719167Z Entering 'third_party/FP16'
2025-12-04T08:53:03.9759322Z Entering 'third_party/FXdiv'
2025-12-04T08:53:03.9797077Z Entering 'third_party/NNPACK'
2025-12-04T08:53:03.9833148Z Entering 'third_party/NVTX'
2025-12-04T08:53:03.9863083Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:03.9903149Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:03.9940870Z Entering 'third_party/aiter'
2025-12-04T08:53:03.9982627Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:04.0015736Z Entering 'third_party/benchmark'
2025-12-04T08:53:04.0039810Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:04.0074579Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:04.0114184Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:04.0151078Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:04.0183077Z Entering 'third_party/cutlass'
2025-12-04T08:53:04.0221497Z Entering 'third_party/fbgemm'
2025-12-04T08:53:04.0250754Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:04.0278850Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:04.0313111Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:04.0334722Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:04.0359092Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:04.0384055Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:04.0411151Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:04.0438340Z Entering 'third_party/flash-attention'
2025-12-04T08:53:04.0473873Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:04.0499996Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:04.0531630Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:04.0557849Z Entering 'third_party/fmt'
2025-12-04T08:53:04.0581168Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:04.0607691Z Entering 'third_party/gloo'
2025-12-04T08:53:04.0631764Z Entering 'third_party/googletest'
2025-12-04T08:53:04.0654041Z Entering 'third_party/ideep'
2025-12-04T08:53:04.0678045Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:04.0703655Z Entering 'third_party/ittapi'
2025-12-04T08:53:04.0726928Z Entering 'third_party/kineto'
2025-12-04T08:53:04.0750345Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:04.0776094Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:04.0807812Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:04.0839016Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:04.0864799Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:04.0889645Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:04.0911363Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:04.0939541Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:04.0962818Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:04.0996152Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:04.1020403Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:04.1046122Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:04.1067948Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:04.1104189Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:04.1134317Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:04.1161901Z Entering 'third_party/kleidiai'
2025-12-04T08:53:04.1191278Z Entering 'third_party/mimalloc'
2025-12-04T08:53:04.1214328Z Entering 'third_party/nlohmann'
2025-12-04T08:53:04.1237808Z Entering 'third_party/onnx'
2025-12-04T08:53:04.1280280Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:04.1307921Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:04.1337713Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:04.1371612Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:04.1400213Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:04.1438886Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:04.1463316Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:04.1484269Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:04.1513273Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:04.1544803Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:04.1582733Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:04.1618147Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:04.1658505Z Entering 'third_party/pocketfft'
2025-12-04T08:53:04.1690096Z Entering 'third_party/protobuf'
2025-12-04T08:53:04.1720064Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:04.1743962Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:04.1769604Z Entering 'third_party/psimd'
2025-12-04T08:53:04.1800645Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:04.1827593Z Entering 'third_party/pybind11'
2025-12-04T08:53:04.1851902Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:04.1890865Z Entering 'third_party/sleef'
2025-12-04T08:53:04.1914719Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:04.1938217Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:04.1963109Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:04.1987805Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:04.2025603Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:04.2052195Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:04.2100131Z [command]/usr/bin/git config --local --name-only --get-regexp http\.https\:\/\/github\.com\/\.extraheader
2025-12-04T08:53:04.2125080Z [command]/usr/bin/git submodule foreach --recursive sh -c "git config --local --name-only --get-regexp 'http\.https\:\/\/github\.com\/\.extraheader' && git config --local --unset-all 'http.https://github.com/.extraheader' || :"
2025-12-04T08:53:04.2297571Z Entering 'android/libs/fbjni'
2025-12-04T08:53:04.2323468Z Entering 'third_party/FP16'
2025-12-04T08:53:04.2347159Z Entering 'third_party/FXdiv'
2025-12-04T08:53:04.2372927Z Entering 'third_party/NNPACK'
2025-12-04T08:53:04.2394271Z Entering 'third_party/NVTX'
2025-12-04T08:53:04.2415646Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:04.2437828Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:04.2468518Z Entering 'third_party/aiter'
2025-12-04T08:53:04.2493403Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:04.2519473Z Entering 'third_party/benchmark'
2025-12-04T08:53:04.2543017Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:04.2568894Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:04.2589562Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:04.2619274Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:04.2643280Z Entering 'third_party/cutlass'
2025-12-04T08:53:04.2668673Z Entering 'third_party/fbgemm'
2025-12-04T08:53:04.2690313Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:04.2711189Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:04.2733763Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:04.2753204Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:04.2783429Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:04.2809275Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:04.2835551Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:04.2861011Z Entering 'third_party/flash-attention'
2025-12-04T08:53:04.2887196Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:04.2922720Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:04.2955607Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:04.2994734Z Entering 'third_party/fmt'
2025-12-04T08:53:04.3021287Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:04.3050490Z Entering 'third_party/gloo'
2025-12-04T08:53:04.3070516Z Entering 'third_party/googletest'
2025-12-04T08:53:04.3098939Z Entering 'third_party/ideep'
2025-12-04T08:53:04.3127079Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:04.3168573Z Entering 'third_party/ittapi'
2025-12-04T08:53:04.3189821Z Entering 'third_party/kineto'
2025-12-04T08:53:04.3218810Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:04.3242528Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:04.3271420Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:04.3294956Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:04.3315021Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:04.3336452Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:04.3360446Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:04.3381779Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:04.3405740Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:04.3430790Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:04.3458430Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:04.3491403Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:04.3513894Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:04.3540017Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:04.3567849Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:04.3592529Z Entering 'third_party/kleidiai'
2025-12-04T08:53:04.3621416Z Entering 'third_party/mimalloc'
2025-12-04T08:53:04.3643296Z Entering 'third_party/nlohmann'
2025-12-04T08:53:04.3678086Z Entering 'third_party/onnx'
2025-12-04T08:53:04.3709940Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:04.3735376Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:04.3759092Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:04.3781928Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:04.3803793Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:04.3824793Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:04.3849808Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:04.3871092Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:04.3896465Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:04.3918754Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:04.3954905Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:04.3980625Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:04.4025691Z Entering 'third_party/pocketfft'
2025-12-04T08:53:04.4051297Z Entering 'third_party/protobuf'
2025-12-04T08:53:04.4074270Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:04.4096114Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:04.4126549Z Entering 'third_party/psimd'
2025-12-04T08:53:04.4152713Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:04.4176901Z Entering 'third_party/pybind11'
2025-12-04T08:53:04.4200361Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:04.4223431Z Entering 'third_party/sleef'
2025-12-04T08:53:04.4256784Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:04.4285895Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:04.4313983Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:04.4341888Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:04.4365666Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:04.4394255Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:04.4434597Z [command]/usr/bin/git config --local --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.4455507Z [command]/usr/bin/git submodule foreach --recursive git config --local --show-origin --name-only --get-regexp remote.origin.url
2025-12-04T08:53:04.4640982Z Entering 'android/libs/fbjni'
2025-12-04T08:53:04.4662819Z file:/home/runner/_work/pytorch/pytorch/.git/modules/android/libs/fbjni/config	remote.origin.url
2025-12-04T08:53:04.4674776Z Entering 'third_party/FP16'
2025-12-04T08:53:04.4689291Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FP16/config	remote.origin.url
2025-12-04T08:53:04.4697121Z Entering 'third_party/FXdiv'
2025-12-04T08:53:04.4709658Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FXdiv/config	remote.origin.url
2025-12-04T08:53:04.4718836Z Entering 'third_party/NNPACK'
2025-12-04T08:53:04.4734889Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK/config	remote.origin.url
2025-12-04T08:53:04.4743962Z Entering 'third_party/NVTX'
2025-12-04T08:53:04.4755149Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NVTX/config	remote.origin.url
2025-12-04T08:53:04.4765358Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:04.4775433Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/VulkanMemoryAllocator/config	remote.origin.url
2025-12-04T08:53:04.4784872Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:04.4795445Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/XNNPACK/config	remote.origin.url
2025-12-04T08:53:04.4810532Z Entering 'third_party/aiter'
2025-12-04T08:53:04.4821583Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/config	remote.origin.url
2025-12-04T08:53:04.4831816Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:04.4840197Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/modules/3rdparty/composable_kernel/config	remote.origin.url
2025-12-04T08:53:04.4855011Z Entering 'third_party/benchmark'
2025-12-04T08:53:04.4865743Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T08:53:04.4874176Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:04.4883624Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/composable_kernel/config	remote.origin.url
2025-12-04T08:53:04.4903906Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:04.4915572Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpp-httplib/config	remote.origin.url
2025-12-04T08:53:04.4927960Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:04.4939288Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpuinfo/config	remote.origin.url
2025-12-04T08:53:04.4947342Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:04.4956609Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cudnn_frontend/config	remote.origin.url
2025-12-04T08:53:04.4965323Z Entering 'third_party/cutlass'
2025-12-04T08:53:04.4974982Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cutlass/config	remote.origin.url
2025-12-04T08:53:04.4994493Z Entering 'third_party/fbgemm'
2025-12-04T08:53:04.5004290Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/config	remote.origin.url
2025-12-04T08:53:04.5014249Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:04.5029229Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/asmjit/config	remote.origin.url
2025-12-04T08:53:04.5038947Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:04.5049173Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/composable_kernel/config	remote.origin.url
2025-12-04T08:53:04.5065373Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:04.5075060Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cpuinfo/config	remote.origin.url
2025-12-04T08:53:04.5083846Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:04.5093809Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cutlass/config	remote.origin.url
2025-12-04T08:53:04.5113218Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:04.5122953Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/googletest/config	remote.origin.url
2025-12-04T08:53:04.5132019Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:04.5141520Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/hipify_torch/config	remote.origin.url
2025-12-04T08:53:04.5151776Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:04.5160906Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/json/config	remote.origin.url
2025-12-04T08:53:04.5172310Z Entering 'third_party/flash-attention'
2025-12-04T08:53:04.5181825Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/config	remote.origin.url
2025-12-04T08:53:04.5190723Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:04.5205142Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/composable_kernel/config	remote.origin.url
2025-12-04T08:53:04.5216393Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:04.5225618Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/cutlass/config	remote.origin.url
2025-12-04T08:53:04.5239973Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:04.5249743Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flatbuffers/config	remote.origin.url
2025-12-04T08:53:04.5260755Z Entering 'third_party/fmt'
2025-12-04T08:53:04.5270618Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fmt/config	remote.origin.url
2025-12-04T08:53:04.5280018Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:04.5290975Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/gemmlowp/gemmlowp/config	remote.origin.url
2025-12-04T08:53:04.5305127Z Entering 'third_party/gloo'
2025-12-04T08:53:04.5315366Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/gloo/config	remote.origin.url
2025-12-04T08:53:04.5324775Z Entering 'third_party/googletest'
2025-12-04T08:53:04.5335696Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:04.5349120Z Entering 'third_party/ideep'
2025-12-04T08:53:04.5362031Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/config	remote.origin.url
2025-12-04T08:53:04.5371085Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:04.5388692Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/modules/mkl-dnn/config	remote.origin.url
2025-12-04T08:53:04.5402665Z Entering 'third_party/ittapi'
2025-12-04T08:53:04.5414035Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ittapi/config	remote.origin.url
2025-12-04T08:53:04.5423104Z Entering 'third_party/kineto'
2025-12-04T08:53:04.5433472Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/config	remote.origin.url
2025-12-04T08:53:04.5442853Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:04.5455124Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/config	remote.origin.url
2025-12-04T08:53:04.5463772Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:04.5474715Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/DCGM/config	remote.origin.url
2025-12-04T08:53:04.5483465Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:04.5493825Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/cpr/config	remote.origin.url
2025-12-04T08:53:04.5502108Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:04.5518146Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/fmt/config	remote.origin.url
2025-12-04T08:53:04.5527180Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:04.5536585Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/config	remote.origin.url
2025-12-04T08:53:04.5545192Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:04.5554587Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/modules/doc/config	remote.origin.url
2025-12-04T08:53:04.5570790Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:04.5579953Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/glog/config	remote.origin.url
2025-12-04T08:53:04.5590448Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:04.5600996Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:04.5609982Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:04.5623327Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/json/config	remote.origin.url
2025-12-04T08:53:04.5633335Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:04.5645011Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/pfs/config	remote.origin.url
2025-12-04T08:53:04.5655628Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:04.5666956Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/config	remote.origin.url
2025-12-04T08:53:04.5678485Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:04.5688441Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/civetweb/config	remote.origin.url
2025-12-04T08:53:04.5698992Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:04.5709491Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/googletest/config	remote.origin.url
2025-12-04T08:53:04.5722476Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:04.5732351Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/fmt/config	remote.origin.url
2025-12-04T08:53:04.5740919Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:04.5750760Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:04.5761873Z Entering 'third_party/kleidiai'
2025-12-04T08:53:04.5771784Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kleidiai/config	remote.origin.url
2025-12-04T08:53:04.5782189Z Entering 'third_party/mimalloc'
2025-12-04T08:53:04.5792204Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/mimalloc/config	remote.origin.url
2025-12-04T08:53:04.5801499Z Entering 'third_party/nlohmann'
2025-12-04T08:53:04.5811113Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/nlohmann/config	remote.origin.url
2025-12-04T08:53:04.5821783Z Entering 'third_party/onnx'
2025-12-04T08:53:04.5837131Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/config	remote.origin.url
2025-12-04T08:53:04.5853600Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:04.5864157Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T08:53:04.5880685Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:04.5893407Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/config	remote.origin.url
2025-12-04T08:53:04.5903568Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:04.5929150Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T08:53:04.5939379Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:04.5950438Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:04.5958223Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:04.5972416Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/ms-gsl/config	remote.origin.url
2025-12-04T08:53:04.5981298Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:04.5996941Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/nlohmann-json/config	remote.origin.url
2025-12-04T08:53:04.6008658Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:04.6025284Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentelemetry-proto/config	remote.origin.url
2025-12-04T08:53:04.6033672Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:04.6049064Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentracing-cpp/config	remote.origin.url
2025-12-04T08:53:04.6060466Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:04.6072009Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/config	remote.origin.url
2025-12-04T08:53:04.6080830Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:04.6100052Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/civetweb/config	remote.origin.url
2025-12-04T08:53:04.6115970Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:04.6126317Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/googletest/config	remote.origin.url
2025-12-04T08:53:04.6138184Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:04.6149178Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/tools/vcpkg/config	remote.origin.url
2025-12-04T08:53:04.6169207Z Entering 'third_party/pocketfft'
2025-12-04T08:53:04.6181392Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/pocketfft/config	remote.origin.url
2025-12-04T08:53:04.6193604Z Entering 'third_party/protobuf'
2025-12-04T08:53:04.6205151Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/config	remote.origin.url
2025-12-04T08:53:04.6216509Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:04.6227858Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T08:53:04.6237550Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:04.6249084Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:04.6259169Z Entering 'third_party/psimd'
2025-12-04T08:53:04.6270026Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/psimd/config	remote.origin.url
2025-12-04T08:53:04.6280350Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:04.6290598Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/pthreadpool/config	remote.origin.url
2025-12-04T08:53:04.6300891Z Entering 'third_party/pybind11'
2025-12-04T08:53:04.6311390Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T08:53:04.6320694Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:04.6330721Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/python-peachpy/config	remote.origin.url
2025-12-04T08:53:04.6344935Z Entering 'third_party/sleef'
2025-12-04T08:53:04.6355168Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/sleef/config	remote.origin.url
2025-12-04T08:53:04.6364161Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:04.6374280Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/config	remote.origin.url
2025-12-04T08:53:04.6388330Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:04.6398230Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:04.6406655Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:04.6416589Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libnop/config	remote.origin.url
2025-12-04T08:53:04.6425213Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:04.6436643Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libuv/config	remote.origin.url
2025-12-04T08:53:04.6445408Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:04.6464252Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T08:53:04.6472202Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:04.6485281Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/modules/tools/clang/config	remote.origin.url
2025-12-04T08:53:04.6511538Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/android/libs/fbjni/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6532375Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FP16/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6549894Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FXdiv/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6568340Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6584738Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NVTX/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6600402Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/VulkanMemoryAllocator/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6614152Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/XNNPACK/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6636096Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6651538Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/modules/3rdparty/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6667026Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/benchmark/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6687071Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6702586Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpp-httplib/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6740904Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpuinfo/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6741515Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cudnn_frontend/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6755209Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cutlass/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6771245Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6785906Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/asmjit/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6803762Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6818419Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cpuinfo/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6834677Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cutlass/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6848909Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6865992Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/hipify_torch/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6880667Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/json/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6895629Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6911971Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6932148Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/cutlass/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6946541Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flatbuffers/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6964848Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fmt/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6981455Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/gemmlowp/gemmlowp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.6995643Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/gloo/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7011324Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7027217Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7042862Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/modules/mkl-dnn/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7061977Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/ittapi/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7077722Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7093351Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7114101Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/DCGM/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7130339Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/cpr/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7146632Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/fmt/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7163811Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7180141Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/modules/doc/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7195289Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/glog/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7217694Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7233166Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/json/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7248684Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/pfs/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7265118Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7281183Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/civetweb/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7300515Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7315565Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/fmt/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7336331Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7351159Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kleidiai/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7364868Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/mimalloc/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7382165Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/nlohmann/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7398687Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7415101Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/modules/third_party/pybind11/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7436781Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7453285Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/benchmark/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7469779Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7485933Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/ms-gsl/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7504820Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/nlohmann-json/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7522508Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentelemetry-proto/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7538646Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentracing-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7554530Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7576662Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/civetweb/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7596923Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7613467Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/tools/vcpkg/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7629254Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/pocketfft/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7645218Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7661903Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/benchmark/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7678770Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7695156Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/psimd/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7710738Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/pthreadpool/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7726713Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/pybind11/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7742812Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/python-peachpy/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7760573Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/sleef/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7776534Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7791970Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7808134Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libnop/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7824021Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libuv/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7839833Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7855260Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/modules/tools/clang/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:04.7872079Z [command]/usr/bin/git config --local http.https://github.com/.extraheader AUTHORIZATION: basic ***
2025-12-04T08:53:04.7899595Z ##[endgroup]
2025-12-04T08:53:04.7899781Z ##[group]Fetching the repository
2025-12-04T08:53:04.7903563Z [command]/usr/bin/git -c protocol.version=2 fetch --prune --no-recurse-submodules origin +refs/heads/*:refs/remotes/origin/* +refs/tags/*:refs/tags/*
2025-12-04T08:53:09.5088149Z From https://github.com/pytorch/pytorch
2025-12-04T08:53:09.5088726Z  * [new branch]                2.6.0.dev20241004+      -> origin/2.6.0.dev20241004+
2025-12-04T08:53:09.5089217Z  * [new branch]                2.9.1                   -> origin/2.9.1
2025-12-04T08:53:09.5089805Z  * [new branch]                AaronWang04_addmmfusion_perftest -> origin/AaronWang04_addmmfusion_perftest
2025-12-04T08:53:09.5090501Z  * [new branch]                Flamefire-patch-1       -> origin/Flamefire-patch-1
2025-12-04T08:53:09.5091092Z  * [new branch]                HDCharles-2.6.0-release-notes -> origin/HDCharles-2.6.0-release-notes
2025-12-04T08:53:09.5091651Z  * [new branch]                HOPrintFunc             -> origin/HOPrintFunc
2025-12-04T08:53:09.5092156Z  * [new branch]                IvanKobzarev/stack/1    -> origin/IvanKobzarev/stack/1
2025-12-04T08:53:09.5092525Z  * [new branch]                NicoshevSVE128          -> origin/NicoshevSVE128
2025-12-04T08:53:09.5092887Z  * [new branch]                PR-AOTInductorNoneBug   -> origin/PR-AOTInductorNoneBug
2025-12-04T08:53:09.5093292Z  * [new branch]                PR-AOTInductorNoneBugFix -> origin/PR-AOTInductorNoneBugFix
2025-12-04T08:53:09.5093688Z  * [new branch]                PR-FixConfigsIssue      -> origin/PR-FixConfigsIssue
2025-12-04T08:53:09.5094057Z  * [new branch]                PR-NoneBugFix-viable    -> origin/PR-NoneBugFix-viable
2025-12-04T08:53:09.5094417Z  * [new branch]                PR-ResetToZero          -> origin/PR-ResetToZero
2025-12-04T08:53:09.5094790Z  * [new branch]                Update-Flash-Packaging  -> origin/Update-Flash-Packaging
2025-12-04T08:53:09.5095158Z  * [new branch]                VLA_exp                 -> origin/VLA_exp
2025-12-04T08:53:09.5095490Z  * [new branch]                activation_bench        -> origin/activation_bench
2025-12-04T08:53:09.5095925Z  * [new branch]                addmm-heuristic         -> origin/addmm-heuristic
2025-12-04T08:53:09.5096270Z  * [new branch]                adi/onednn_aarch64      -> origin/adi/onednn_aarch64
2025-12-04T08:53:09.5096608Z  * [new branch]                adi/test                -> origin/adi/test
2025-12-04T08:53:09.5096933Z  * [new branch]                adi/test_bgemm          -> origin/adi/test_bgemm
2025-12-04T08:53:09.5097258Z  * [new branch]                adi/test_m8g            -> origin/adi/test_m8g
2025-12-04T08:53:09.5097594Z  * [new branch]                adi/test_onednn         -> origin/adi/test_onednn
2025-12-04T08:53:09.5097946Z  * [new branch]                adi/test_onednn_v3.9    -> origin/adi/test_onednn_v3.9
2025-12-04T08:53:09.5098310Z  * [new branch]                adi/test_presve_change  -> origin/adi/test_presve_change
2025-12-04T08:53:09.5099358Z  * [new branch]                adi/test_timm           -> origin/adi/test_timm
2025-12-04T08:53:09.5099831Z  * [new branch]                adi/testpresve_change   -> origin/adi/testpresve_change
2025-12-04T08:53:09.5100210Z  * [new branch]                aditew01/test/vec_bf16  -> origin/aditew01/test/vec_bf16
2025-12-04T08:53:09.5100621Z  * [new branch]                ah-globalfeedback-hook  -> origin/ah-globalfeedback-hook
2025-12-04T08:53:09.5101024Z  * [new branch]                albanD-patch-1          -> origin/albanD-patch-1
2025-12-04T08:53:09.5101385Z  * [new branch]                also-surround-shimh     -> origin/also-surround-shimh
2025-12-04T08:53:09.5101753Z  * [new branch]                angelayi/aot_compile    -> origin/angelayi/aot_compile
2025-12-04T08:53:09.5102173Z  * [new branch]                angelayi/aoti_additional_files -> origin/angelayi/aoti_additional_files
2025-12-04T08:53:09.5102468Z  * [new branch]                angelayi/benchmark      -> origin/angelayi/benchmark
2025-12-04T08:53:09.5102802Z  * [new branch]                angelayi/change_pytree_serialization -> origin/angelayi/change_pytree_serialization
2025-12-04T08:53:09.5103122Z  * [new branch]                angelayi/cpp_loader     -> origin/angelayi/cpp_loader
2025-12-04T08:53:09.5103398Z  * [new branch]                angelayi/inductor_const -> origin/angelayi/inductor_const
2025-12-04T08:53:09.5103664Z  * [new branch]                angelayi/lstm           -> origin/angelayi/lstm
2025-12-04T08:53:09.5103916Z  * [new branch]                angelayi/no_so_weight   -> origin/angelayi/no_so_weight
2025-12-04T08:53:09.5104188Z  * [new branch]                angelayi/scan_layers    -> origin/angelayi/scan_layers
2025-12-04T08:53:09.5104444Z  * [new branch]                angelayi/side_eff       -> origin/angelayi/side_eff
2025-12-04T08:53:09.5104693Z  * [new branch]                angelayi/state_dict     -> origin/angelayi/state_dict
2025-12-04T08:53:09.5104954Z  * [new branch]                angelayi/symint_input   -> origin/angelayi/symint_input
2025-12-04T08:53:09.5105217Z  * [new branch]                angelayi/symm_mem       -> origin/angelayi/symm_mem
2025-12-04T08:53:09.5105469Z  * [new branch]                angelayi/test_cpp       -> origin/angelayi/test_cpp
2025-12-04T08:53:09.5105722Z  * [new branch]                angelayi/torch_size     -> origin/angelayi/torch_size
2025-12-04T08:53:09.5105972Z  * [new branch]                annotate_assert         -> origin/annotate_assert
2025-12-04T08:53:09.5106234Z  * [new branch]                annotate_fallback_kernel -> origin/annotate_fallback_kernel
2025-12-04T08:53:09.5106501Z  * [new branch]                annotation_deepcopy     -> origin/annotation_deepcopy
2025-12-04T08:53:09.5106756Z  * [new branch]                annotation_dynamo       -> origin/annotation_dynamo
2025-12-04T08:53:09.5107008Z  * [new branch]                aot_eager_stack_trace   -> origin/aot_eager_stack_trace
2025-12-04T08:53:09.5107261Z  * [new branch]                aoti-cuda-alloc         -> origin/aoti-cuda-alloc
2025-12-04T08:53:09.5107522Z  * [new branch]                aoti_const_device       -> origin/aoti_const_device
2025-12-04T08:53:09.5107778Z  * [new branch]                aoti_fqn_name_interface -> origin/aoti_fqn_name_interface
2025-12-04T08:53:09.5108067Z  * [new branch]                aoti_package_weights_binary -> origin/aoti_package_weights_binary
2025-12-04T08:53:09.5108344Z  * [new branch]                aoti_target_windows     -> origin/aoti_target_windows
2025-12-04T08:53:09.5108650Z  * [new branch]                arsh/feat/inductor_check_profiling -> origin/arsh/feat/inductor_check_profiling
2025-12-04T08:53:09.5108953Z  * [new branch]                async_tp                -> origin/async_tp
2025-12-04T08:53:09.5109226Z  * [new branch]                atalman-inductor-perf-cu124 -> origin/atalman-inductor-perf-cu124
2025-12-04T08:53:09.5109573Z  * [new branch]                atalman-inductor-perf-cu124.1 -> origin/atalman-inductor-perf-cu124.1
2025-12-04T08:53:09.5110012Z  * [new branch]                atalman-patch-2         -> origin/atalman-patch-2
2025-12-04T08:53:09.5110257Z  * [new branch]                atalman-patch-3         -> origin/atalman-patch-3
2025-12-04T08:53:09.5110593Z  * [new branch]                atalman-patch-4         -> origin/atalman-patch-4
2025-12-04T08:53:09.5110844Z  * [new branch]                atalman-patch-5         -> origin/atalman-patch-5
2025-12-04T08:53:09.5111084Z  * [new branch]                atalman-patch-6         -> origin/atalman-patch-6
2025-12-04T08:53:09.5111327Z  * [new branch]                atalman-patch-7         -> origin/atalman-patch-7
2025-12-04T08:53:09.5111572Z  * [new branch]                atalman-patch-8         -> origin/atalman-patch-8
2025-12-04T08:53:09.5111826Z  * [new branch]                atalman_inductor_2.3.1  -> origin/atalman_inductor_2.3.1
2025-12-04T08:53:09.5112138Z  * [new branch]                atalman_inductor_2.4.0  -> origin/atalman_inductor_2.4.0
2025-12-04T08:53:09.5112394Z  * [new branch]                atalman_inductor_2.4.x  -> origin/atalman_inductor_2.4.x
2025-12-04T08:53:09.5112624Z  * [new branch]                attention_benchmarking_clean -> origin/attention_benchmarking_clean
2025-12-04T08:53:09.5112867Z  * [new branch]                bahuang/dt_fix_scalar_add -> origin/bahuang/dt_fix_scalar_add
2025-12-04T08:53:09.5113087Z  * [new branch]                bahuang/fix_debug_mode  -> origin/bahuang/fix_debug_mode
2025-12-04T08:53:09.5113293Z  * [new branch]                bahuang/fix_expand      -> origin/bahuang/fix_expand
2025-12-04T08:53:09.5113495Z  * [new branch]                bahuang/test            -> origin/bahuang/test
2025-12-04T08:53:09.5113676Z  * [new branch]                base/1.5                -> origin/base/1.5
2025-12-04T08:53:09.5113900Z  * [new branch]                batching_sdpa_efficient_attention -> origin/batching_sdpa_efficient_attention
2025-12-04T08:53:09.5114143Z  * [new branch]                bench_scaled_mm_ops     -> origin/bench_scaled_mm_ops
2025-12-04T08:53:09.5114349Z  * [new branch]                benchmark-updates       -> origin/benchmark-updates
2025-12-04T08:53:09.5114596Z  * [new branch]                benchmarking-script     -> origin/benchmarking-script
2025-12-04T08:53:09.5114808Z  * [new branch]                bertmaher/pinbump26     -> origin/bertmaher/pinbump26
2025-12-04T08:53:09.5115008Z  * [new branch]                bertrand/cutlass        -> origin/bertrand/cutlass
2025-12-04T08:53:09.5115212Z  * [new branch]                bf/bug-static-input     -> origin/bf/bug-static-input
2025-12-04T08:53:09.5115417Z  * [new branch]                bf/cg-backend           -> origin/bf/cg-backend
2025-12-04T08:53:09.5115606Z  * [new branch]                bf/cg-nccl-test         -> origin/bf/cg-nccl-test
2025-12-04T08:53:09.5115800Z  * [new branch]                bf/cg-remove-check      -> origin/bf/cg-remove-check
2025-12-04T08:53:09.5116013Z  * [new branch]                bf/clean-torchbench-hf  -> origin/bf/clean-torchbench-hf
2025-12-04T08:53:09.5116225Z  * [new branch]                bf/combo-debug-log      -> origin/bf/combo-debug-log
2025-12-04T08:53:09.5116417Z  * [new branch]                bf/cudagraph            -> origin/bf/cudagraph
2025-12-04T08:53:09.5116680Z  * [new branch]                bf/cudagraph-disable-input-mutation -> origin/bf/cudagraph-disable-input-mutation
2025-12-04T08:53:09.5117064Z  * [new branch]                bf/cudagraph-enable-input-mutation-support-benchmark -> origin/bf/cudagraph-enable-input-mutation-support-benchmark
2025-12-04T08:53:09.5117416Z  * [new branch]                bf/cudagraph-partition  -> origin/bf/cudagraph-partition
2025-12-04T08:53:09.5117652Z  * [new branch]                bf/donated-buffer-bench -> origin/bf/donated-buffer-bench
2025-12-04T08:53:09.5117865Z  * [new branch]                bf/dynamo-partition     -> origin/bf/dynamo-partition
2025-12-04T08:53:09.5118062Z  * [new branch]                bf/lite                 -> origin/bf/lite
2025-12-04T08:53:09.5118368Z  * [new branch]                bf/pa-non-divisible     -> origin/bf/pa-non-divisible
2025-12-04T08:53:09.5118607Z  * [new branch]                bf/partition-cache-free-symbols -> origin/bf/partition-cache-free-symbols
2025-12-04T08:53:09.5118917Z  * [new branch]                bf/partition-memory-plan -> origin/bf/partition-memory-plan
2025-12-04T08:53:09.5119153Z  * [new branch]                bf/partition-move-cpu   -> origin/bf/partition-move-cpu
2025-12-04T08:53:09.5119386Z  * [new branch]                bf/partition-view-fallback -> origin/bf/partition-view-fallback
2025-12-04T08:53:09.5119633Z  * [new branch]                bf/remove-check-55b0c39d -> origin/bf/remove-check-55b0c39d
2025-12-04T08:53:09.5119848Z  * [new branch]                bf/timm-nov-26-2025     -> origin/bf/timm-nov-26-2025
2025-12-04T08:53:09.5120073Z  * [new branch]                bf/transformer-pin-4-57-3 -> origin/bf/transformer-pin-4-57-3
2025-12-04T08:53:09.5120321Z  * [new branch]                bisect_perf_hf_T5_3acc6eac492 -> origin/bisect_perf_hf_T5_3acc6eac492
2025-12-04T08:53:09.5120601Z  * [new branch]                bisect_perf_hf_T5_3fcf66f61fb -> origin/bisect_perf_hf_T5_3fcf66f61fb
2025-12-04T08:53:09.5120843Z  * [new branch]                bisect_perf_hf_T5_4009d154129 -> origin/bisect_perf_hf_T5_4009d154129
2025-12-04T08:53:09.5121082Z  * [new branch]                bisect_perf_hf_T5_40d0740e73d -> origin/bisect_perf_hf_T5_40d0740e73d
2025-12-04T08:53:09.5121307Z  * [new branch]                bisect_perf_hf_T5_5268754e -> origin/bisect_perf_hf_T5_5268754e
2025-12-04T08:53:09.5121537Z  * [new branch]                bisect_perf_hf_T5_7d89a8d385c -> origin/bisect_perf_hf_T5_7d89a8d385c
2025-12-04T08:53:09.5121779Z  * [new branch]                bisect_perf_hf_T5_b7a25c1ee7c -> origin/bisect_perf_hf_T5_b7a25c1ee7c
2025-12-04T08:53:09.5122010Z  * [new branch]                bisect_perf_hf_T5_c25b201583f -> origin/bisect_perf_hf_T5_c25b201583f
2025-12-04T08:53:09.5122220Z  * [new branch]                bisect_perf_hf_T5_c93e57efac0 -> origin/bisect_perf_hf_T5_c93e57efac0
2025-12-04T08:53:09.5122432Z  * [new branch]                bisect_perf_hf_T5_ca9813ea149 -> origin/bisect_perf_hf_T5_ca9813ea149
2025-12-04T08:53:09.5122644Z  * [new branch]                bisect_perf_hf_T5_d65f194a -> origin/bisect_perf_hf_T5_d65f194a
2025-12-04T08:53:09.5122849Z  * [new branch]                bisect_perf_hf_T5_da94ab0b -> origin/bisect_perf_hf_T5_da94ab0b
2025-12-04T08:53:09.5123059Z  * [new branch]                bisect_perf_hf_T5_da94ab0b_new -> origin/bisect_perf_hf_T5_da94ab0b_new
2025-12-04T08:53:09.5123271Z  * [new branch]                bisect_perf_hf_T5_db4e8a1d8a8 -> origin/bisect_perf_hf_T5_db4e8a1d8a8
2025-12-04T08:53:09.5123484Z  * [new branch]                bisect_perf_hf_T5_e0d97e936a2 -> origin/bisect_perf_hf_T5_e0d97e936a2
2025-12-04T08:53:09.5123702Z  * [new branch]                bisect_perf_hf_T5_f23621ec563 -> origin/bisect_perf_hf_T5_f23621ec563
2025-12-04T08:53:09.5123906Z  * [new branch]                brister/fx_device_type  -> origin/brister/fx_device_type
2025-12-04T08:53:09.5124129Z  * [new branch]                brister/test_inductor_all_fx -> origin/brister/test_inductor_all_fx
2025-12-04T08:53:09.5124390Z  * [new branch]                brister/tiled_reduction_no_numel_check -> origin/brister/tiled_reduction_no_numel_check
2025-12-04T08:53:09.5124612Z  * [new branch]                bwd-backup              -> origin/bwd-backup
2025-12-04T08:53:09.5124785Z  * [new branch]                c57382a49               -> origin/c57382a49
2025-12-04T08:53:09.5124952Z  * [new branch]                ca_0431d47eaa           -> origin/ca_0431d47eaa
2025-12-04T08:53:09.5125120Z  * [new branch]                ca_fix_0431d47eaa       -> origin/ca_fix_0431d47eaa
2025-12-04T08:53:09.5125325Z  * [new branch]                camyllh/test_setup_hooks_push -> origin/camyllh/test_setup_hooks_push
2025-12-04T08:53:09.5125541Z  * [new branch]                cccclai-patch-1         -> origin/cccclai-patch-1
2025-12-04T08:53:09.5126364Z  * [new branch]                cherry-pick-159969-by-pytorch_bot_bot_ -> origin/cherry-pick-159969-by-pytorch_bot_bot_
2025-12-04T08:53:09.5126684Z  * [new branch]                cherry-pick-160586-by-pytorch_bot_bot_ -> origin/cherry-pick-160586-by-pytorch_bot_bot_
2025-12-04T08:53:09.5126963Z  * [new branch]                cherry-pick-162208-by-pytorch_bot_bot_ -> origin/cherry-pick-162208-by-pytorch_bot_bot_
2025-12-04T08:53:09.5127243Z  * [new branch]                cherry-pick-163169-by-pytorch_bot_bot_ -> origin/cherry-pick-163169-by-pytorch_bot_bot_
2025-12-04T08:53:09.5127518Z  * [new branch]                cherry-pick-165086-by-pytorch_bot_bot_ -> origin/cherry-pick-165086-by-pytorch_bot_bot_
2025-12-04T08:53:09.5127794Z  * [new branch]                cherry-pick-165514-by-pytorch_bot_bot_ -> origin/cherry-pick-165514-by-pytorch_bot_bot_
2025-12-04T08:53:09.5128070Z  * [new branch]                cherry-pick-165601-by-pytorch_bot_bot_ -> origin/cherry-pick-165601-by-pytorch_bot_bot_
2025-12-04T08:53:09.5128352Z  * [new branch]                cherry-pick-165667-by-pytorch_bot_bot_ -> origin/cherry-pick-165667-by-pytorch_bot_bot_
2025-12-04T08:53:09.5128630Z  * [new branch]                cherry-pick-165815-by-pytorch_bot_bot_ -> origin/cherry-pick-165815-by-pytorch_bot_bot_
2025-12-04T08:53:09.5128905Z  * [new branch]                cherry-pick-165922-by-pytorch_bot_bot_ -> origin/cherry-pick-165922-by-pytorch_bot_bot_
2025-12-04T08:53:09.5129182Z  * [new branch]                cherry-pick-166148-by-pytorch_bot_bot_ -> origin/cherry-pick-166148-by-pytorch_bot_bot_
2025-12-04T08:53:09.5129457Z  * [new branch]                cherry-pick-166181-by-pytorch_bot_bot_ -> origin/cherry-pick-166181-by-pytorch_bot_bot_
2025-12-04T08:53:09.5129733Z  * [new branch]                cherry-pick-166404-by-pytorch_bot_bot_ -> origin/cherry-pick-166404-by-pytorch_bot_bot_
2025-12-04T08:53:09.5130009Z  * [new branch]                cherry-pick-166427-by-pytorch_bot_bot_ -> origin/cherry-pick-166427-by-pytorch_bot_bot_
2025-12-04T08:53:09.5130288Z  * [new branch]                cherry-pick-166480-by-pytorch_bot_bot_ -> origin/cherry-pick-166480-by-pytorch_bot_bot_
2025-12-04T08:53:09.5130623Z  * [new branch]                cherry-pick-166570-by-pytorch_bot_bot_ -> origin/cherry-pick-166570-by-pytorch_bot_bot_
2025-12-04T08:53:09.5130899Z  * [new branch]                cherry-pick-166993-by-pytorch_bot_bot_ -> origin/cherry-pick-166993-by-pytorch_bot_bot_
2025-12-04T08:53:09.5131176Z  * [new branch]                cherry-pick-167111-by-pytorch_bot_bot_ -> origin/cherry-pick-167111-by-pytorch_bot_bot_
2025-12-04T08:53:09.5131456Z  * [new branch]                cherry-pick-167478-by-pytorch_bot_bot_ -> origin/cherry-pick-167478-by-pytorch_bot_bot_
2025-12-04T08:53:09.5131695Z  * [new branch]                cherry_pick_166036_166040 -> origin/cherry_pick_166036_166040
2025-12-04T08:53:09.5131891Z  * [new branch]                cherry_pick_166457      -> origin/cherry_pick_166457
2025-12-04T08:53:09.5132072Z  * [new branch]                cherrypick_166338       -> origin/cherrypick_166338
2025-12-04T08:53:09.5132256Z  * [new branch]                cherrypick_166458       -> origin/cherrypick_166458
2025-12-04T08:53:09.5132444Z  * [new branch]                cherrypick_166586       -> origin/cherrypick_166586
2025-12-04T08:53:09.5132619Z  * [new branch]                cherrypick_166956       -> origin/cherrypick_166956
2025-12-04T08:53:09.5132794Z  * [new branch]                ci_attn                 -> origin/ci_attn
2025-12-04T08:53:09.5132965Z  * [new branch]                codex-testing           -> origin/codex-testing
2025-12-04T08:53:09.5133226Z  * [new branch]                codex/add-check_memory_overlap-helper-functions -> origin/codex/add-check_memory_overlap-helper-functions
2025-12-04T08:53:09.5133537Z  * [new branch]                codex/fix-issue-121219-in-pytorch -> origin/codex/fix-issue-121219-in-pytorch
2025-12-04T08:53:09.5133858Z  * [new branch]                codex/investigate-segfaults-in-get_tensor_storage_id -> origin/codex/investigate-segfaults-in-get_tensor_storage_id
2025-12-04T08:53:09.5134274Z  * [new branch]                codex/refactor-lintrunner-config-to-use-uv-run -> origin/codex/refactor-lintrunner-config-to-use-uv-run
2025-12-04T08:53:09.5134579Z  * [new branch]                compatiblpy39util       -> origin/compatiblpy39util
2025-12-04T08:53:09.5134762Z  * [new branch]                cond_hop_device         -> origin/cond_hop_device
2025-12-04T08:53:09.5134933Z  * [new branch]                context_test            -> origin/context_test
2025-12-04T08:53:09.5135196Z  * [new branch]                copilot/code-style-cleanup-python-pip -> origin/copilot/code-style-cleanup-python-pip
2025-12-04T08:53:09.5135436Z  * [new branch]                cpio/fix_new_ami_tests  -> origin/cpio/fix_new_ami_tests
2025-12-04T08:53:09.5135658Z  * [new branch]                cpp-docs-dependency-upgrade -> origin/cpp-docs-dependency-upgrade
2025-12-04T08:53:09.5135875Z  * [new branch]                csl/always_produce_xml  -> origin/csl/always_produce_xml
2025-12-04T08:53:09.5136085Z  * [new branch]                csl/build_test_more_procs -> origin/csl/build_test_more_procs
2025-12-04T08:53:09.5136295Z  * [new branch]                csl/build_test_more_procs2 -> origin/csl/build_test_more_procs2
2025-12-04T08:53:09.5136484Z  * [new branch]                csl/clean_up            -> origin/csl/clean_up
2025-12-04T08:53:09.5136672Z  * [new branch]                csl/fix_retry_segfault_exit -> origin/csl/fix_retry_segfault_exit
2025-12-04T08:53:09.5136866Z  * [new branch]                csl/katex               -> origin/csl/katex
2025-12-04T08:53:09.5137032Z  * [new branch]                csl/larger_runner       -> origin/csl/larger_runner
2025-12-04T08:53:09.5137207Z  * [new branch]                csl/lint_testing        -> origin/csl/lint_testing
2025-12-04T08:53:09.5137379Z  * [new branch]                csl/lint_thing          -> origin/csl/lint_thing
2025-12-04T08:53:09.5137558Z  * [new branch]                csl/lintrunner_stuff    -> origin/csl/lintrunner_stuff
2025-12-04T08:53:09.5137764Z  * [new branch]                csl/manually_gen_json   -> origin/csl/manually_gen_json
2025-12-04T08:53:09.5137948Z  * [new branch]                csl/mps_sharding        -> origin/csl/mps_sharding
2025-12-04T08:53:09.5138129Z  * [new branch]                csl/multistage_docker   -> origin/csl/multistage_docker
2025-12-04T08:53:09.5138311Z  * [new branch]                csl/print_timing        -> origin/csl/print_timing
2025-12-04T08:53:09.5138493Z  * [new branch]                csl/remove_experiment   -> origin/csl/remove_experiment
2025-12-04T08:53:09.5138701Z  * [new branch]                csl/remove_maybe_unused_var -> origin/csl/remove_maybe_unused_var
2025-12-04T08:53:09.5138932Z  * [new branch]                csl/remove_repo_specific_autolabel -> origin/csl/remove_repo_specific_autolabel
2025-12-04T08:53:09.5139154Z  * [new branch]                csl/remove_run_parallel -> origin/csl/remove_run_parallel
2025-12-04T08:53:09.5139347Z  * [new branch]                csl/remove_unused_vars  -> origin/csl/remove_unused_vars
2025-12-04T08:53:09.5139529Z  * [new branch]                csl/revert_open         -> origin/csl/revert_open
2025-12-04T08:53:09.5139709Z  * [new branch]                csl/skip_build          -> origin/csl/skip_build
2025-12-04T08:53:09.5139901Z  * [new branch]                csl/smaller_avx_amx_runenrs -> origin/csl/smaller_avx_amx_runenrs
2025-12-04T08:53:09.5140094Z  * [new branch]                csl/td_job_level        -> origin/csl/td_job_level
2025-12-04T08:53:09.5140297Z  * [new branch]                csl/test_cuda_build_large_runner -> origin/csl/test_cuda_build_large_runner
2025-12-04T08:53:09.5140566Z  * [new branch]                csl/test_owners_autograd_dispatch_nn -> origin/csl/test_owners_autograd_dispatch_nn
2025-12-04T08:53:09.5140814Z  * [new branch]                csl/test_owners_higher_confidence -> origin/csl/test_owners_higher_confidence
2025-12-04T08:53:09.5141032Z  * [new branch]                csl/upload_json_running -> origin/csl/upload_json_running
2025-12-04T08:53:09.5141269Z  * [new branch]                csl/win_sccache         -> origin/csl/win_sccache
2025-12-04T08:53:09.5141471Z  * [new branch]                csl/xml_stuff           -> origin/csl/xml_stuff
2025-12-04T08:53:09.5141640Z  * [new branch]                cublasrelax2            -> origin/cublasrelax2
2025-12-04T08:53:09.5141806Z  * [new branch]                cuda_mempool            -> origin/cuda_mempool
2025-12-04T08:53:09.5141982Z  * [new branch]                custom_lowering_dict    -> origin/custom_lowering_dict
2025-12-04T08:53:09.5142176Z  * [new branch]                d4l3k/debug_plane_frtrace -> origin/d4l3k/debug_plane_frtrace
2025-12-04T08:53:09.5142365Z  * [new branch]                daxia6/2.8o3            -> origin/daxia6/2.8o3
2025-12-04T08:53:09.5142532Z  * [new branch]                debug-guard             -> origin/debug-guard
2025-12-04T08:53:09.5142706Z  * [new branch]                delete-quant-docs       -> origin/delete-quant-docs
2025-12-04T08:53:09.5143044Z  * [new branch]                dependabot/pip/dot-ci/docker/ci_commit_pins/main/transformers-4.57.0 -> origin/dependabot/pip/dot-ci/docker/ci_commit_pins/main/transformers-4.57.0
2025-12-04T08:53:09.5143509Z  * [new branch]                dependabot/pip/dot-ci/docker/ci_commit_pins/main/transformers-4.57.1 -> origin/dependabot/pip/dot-ci/docker/ci_commit_pins/main/transformers-4.57.1
2025-12-04T08:53:09.5144014Z  * [new branch]                desertfire/test_cpp_wrapper -> origin/desertfire/test_cpp_wrapper
2025-12-04T08:53:09.5144252Z  * [new branch]                desertfire/triton-cpu-for-aarch64 -> origin/desertfire/triton-cpu-for-aarch64
2025-12-04T08:53:09.5144488Z  * [new branch]                dev/dhruva/flex_attn_opt -> origin/dev/dhruva/flex_attn_opt
2025-12-04T08:53:09.5144688Z  * [new branch]                dev/joona/MPSNDArrayAdd -> origin/dev/joona/MPSNDArrayAdd
2025-12-04T08:53:09.5144883Z  * [new branch]                dev/joona/Unranked      -> origin/dev/joona/Unranked
2025-12-04T08:53:09.5145063Z  * [new branch]                dev/joona/cat           -> origin/dev/joona/cat
2025-12-04T08:53:09.5145244Z  * [new branch]                dev/joona/embeddingbag  -> origin/dev/joona/embeddingbag
2025-12-04T08:53:09.5145453Z  * [new branch]                dev/joona/fix_sdpa_memtest -> origin/dev/joona/fix_sdpa_memtest
2025-12-04T08:53:09.5145671Z  * [new branch]                dev/joona/getTensorsString -> origin/dev/joona/getTensorsString
2025-12-04T08:53:09.5145892Z  * [new branch]                dev/joona/mps_linear_macos14 -> origin/dev/joona/mps_linear_macos14
2025-12-04T08:53:09.5146100Z  * [new branch]                dev/joona/scalar_clamp  -> origin/dev/joona/scalar_clamp
2025-12-04T08:53:09.5146288Z  * [new branch]                dev/joona/sdpa          -> origin/dev/joona/sdpa
2025-12-04T08:53:09.5146466Z  * [new branch]                dev/joona/sdpa_api      -> origin/dev/joona/sdpa_api
2025-12-04T08:53:09.5146648Z  * [new branch]                dev/joona/type_inf      -> origin/dev/joona/type_inf
2025-12-04T08:53:09.5146846Z  * [new branch]                dev/joona/ulpAssertClose -> origin/dev/joona/ulpAssertClose
2025-12-04T08:53:09.5147043Z  * [new branch]                dev/joona/upsize3d      -> origin/dev/joona/upsize3d
2025-12-04T08:53:09.5147227Z  * [new branch]                disp_counter            -> origin/disp_counter
2025-12-04T08:53:09.5147405Z  * [new branch]                divyanshk-patch-1       -> origin/divyanshk-patch-1
2025-12-04T08:53:09.5147580Z  * [new branch]                docs                    -> origin/docs
2025-12-04T08:53:09.5147744Z  * [new branch]                documentation           -> origin/documentation
2025-12-04T08:53:09.5147924Z  * [new branch]                eager_model_benchmarks  -> origin/eager_model_benchmarks
2025-12-04T08:53:09.5148129Z  * [new branch]                embg/test_inductor_ci_control -> origin/embg/test_inductor_ci_control
2025-12-04T08:53:09.5148356Z  * [new branch]                embg/triton_l2_prefetch_128B -> origin/embg/triton_l2_prefetch_128B
2025-12-04T08:53:09.5148610Z  * [new branch]                embg/triton_l2_prefetch_256B -> origin/embg/triton_l2_prefetch_256B
2025-12-04T08:53:09.5148826Z  * [new branch]                eqy-patch-1             -> origin/eqy-patch-1
2025-12-04T08:53:09.5148996Z  * [new branch]                eqy-patch-2             -> origin/eqy-patch-2
2025-12-04T08:53:09.5149163Z  * [new branch]                eqy-patch-3             -> origin/eqy-patch-3
2025-12-04T08:53:09.5149330Z  * [new branch]                eqy-patch-4             -> origin/eqy-patch-4
2025-12-04T08:53:09.5149494Z  * [new branch]                eqy-patch-5             -> origin/eqy-patch-5
2025-12-04T08:53:09.5149656Z  * [new branch]                eqy-patch-6             -> origin/eqy-patch-6
2025-12-04T08:53:09.5149835Z  * [new branch]                exclamaforte/amd-ma     -> origin/exclamaforte/amd-ma
2025-12-04T08:53:09.5150067Z  * [new branch]                exclamaforte/combo-kernels-perf-run -> origin/exclamaforte/combo-kernels-perf-run
2025-12-04T08:53:09.5150326Z  * [new branch]                exclamaforte/do_bench_refactor -> origin/exclamaforte/do_bench_refactor
2025-12-04T08:53:09.5150630Z  * [new branch]                exclamaforte/enable-mem-dep-fusion -> origin/exclamaforte/enable-mem-dep-fusion
2025-12-04T08:53:09.5150916Z  * [new branch]                exclamaforte/fix-exhaustive-autotuning -> origin/exclamaforte/fix-exhaustive-autotuning
2025-12-04T08:53:09.5151207Z  * [new branch]                exclamaforte/fix-trace-parsing-fx-svg -> origin/exclamaforte/fix-trace-parsing-fx-svg
2025-12-04T08:53:09.5151510Z  * [new branch]                exclamaforte/force-pointwise-cat-perf-run -> origin/exclamaforte/force-pointwise-cat-perf-run
2025-12-04T08:53:09.5151772Z  * [new branch]                exclamaforte/fusion-data -> origin/exclamaforte/fusion-data
2025-12-04T08:53:09.5152006Z  * [new branch]                exclamaforte/gemm-benchmark-run -> origin/exclamaforte/gemm-benchmark-run
2025-12-04T08:53:09.5152253Z  * [new branch]                exclamaforte/gemm-export-model -> origin/exclamaforte/gemm-export-model
2025-12-04T08:53:09.5152475Z  * [new branch]                exclamaforte/gemm-model -> origin/exclamaforte/gemm-model
2025-12-04T08:53:09.5152737Z  * [new branch]                exclamaforte/gemm-model-all-data-collection -> origin/exclamaforte/gemm-model-all-data-collection
2025-12-04T08:53:09.5153007Z  * [new branch]                exclamaforte/gemm-to-amd -> origin/exclamaforte/gemm-to-amd
2025-12-04T08:53:09.5153227Z  * [new branch]                exclamaforte/just-gemm-model -> origin/exclamaforte/just-gemm-model
2025-12-04T08:53:09.5153492Z  * [new branch]                exclamaforte/just-gemm-model-no-refactor -> origin/exclamaforte/just-gemm-model-no-refactor
2025-12-04T08:53:09.5153767Z  * [new branch]                exclamaforte/profile-diff-algo -> origin/exclamaforte/profile-diff-algo
2025-12-04T08:53:09.5154027Z  * [new branch]                exclamaforte/profiler-visualization -> origin/exclamaforte/profiler-visualization
2025-12-04T08:53:09.5154294Z  * [new branch]                exclamaforte/test_cpp_wrapper_mode -> origin/exclamaforte/test_cpp_wrapper_mode
2025-12-04T08:53:09.5154568Z  * [new branch]                exclamaforte/update-autotune-configs -> origin/exclamaforte/update-autotune-configs
2025-12-04T08:53:09.5154859Z  * [new branch]                exclamaforte/update-autotune-configs-2 -> origin/exclamaforte/update-autotune-configs-2
2025-12-04T08:53:09.5155092Z  * [new branch]                exec                    -> origin/exec
2025-12-04T08:53:09.5155264Z  * [new branch]                experimental-mosaic     -> origin/experimental-mosaic
2025-12-04T08:53:09.5155450Z  * [new branch]                export-D61047529        -> origin/export-D61047529
2025-12-04T08:53:09.5155623Z  * [new branch]                export-D71412006        -> origin/export-D71412006
2025-12-04T08:53:09.5155857Z  * [new branch]                export-D73042989        -> origin/export-D73042989
2025-12-04T08:53:09.5156088Z  * [new branch]                export-D78957093        -> origin/export-D78957093
2025-12-04T08:53:09.5156284Z  * [new branch]                export-D78996107        -> origin/export-D78996107
2025-12-04T08:53:09.5156457Z  * [new branch]                export-D80823877        -> origin/export-D80823877
2025-12-04T08:53:09.5156630Z  * [new branch]                export-D80958642        -> origin/export-D80958642
2025-12-04T08:53:09.5156805Z  * [new branch]                export-D81054193        -> origin/export-D81054193
2025-12-04T08:53:09.5156973Z  * [new branch]                export-D81204584        -> origin/export-D81204584
2025-12-04T08:53:09.5157144Z  * [new branch]                export-D81429090        -> origin/export-D81429090
2025-12-04T08:53:09.5157312Z  * [new branch]                export-D82250826        -> origin/export-D82250826
2025-12-04T08:53:09.5157488Z  * [new branch]                export-D82253817        -> origin/export-D82253817
2025-12-04T08:53:09.5157661Z  * [new branch]                export-D83541846        -> origin/export-D83541846
2025-12-04T08:53:09.5157828Z  * [new branch]                export-D83627170        -> origin/export-D83627170
2025-12-04T08:53:09.5157998Z  * [new branch]                export-D83766701        -> origin/export-D83766701
2025-12-04T08:53:09.5158168Z  * [new branch]                export-D83768878        -> origin/export-D83768878
2025-12-04T08:53:09.5158343Z  * [new branch]                export-D83769447        -> origin/export-D83769447
2025-12-04T08:53:09.5158513Z  * [new branch]                export-D84089824        -> origin/export-D84089824
2025-12-04T08:53:09.5158683Z  * [new branch]                export-D84213020        -> origin/export-D84213020
2025-12-04T08:53:09.5158850Z  * [new branch]                export-D84373821        -> origin/export-D84373821
2025-12-04T08:53:09.5159020Z  * [new branch]                export-D84612194        -> origin/export-D84612194
2025-12-04T08:53:09.5159198Z  * [new branch]                export-D84890985        -> origin/export-D84890985
2025-12-04T08:53:09.5159492Z  * [new branch]                export-D85122326        -> origin/export-D85122326
2025-12-04T08:53:09.5159666Z  * [new branch]                export-D86256198        -> origin/export-D86256198
2025-12-04T08:53:09.5159834Z  * [new branch]                export-D86460608        -> origin/export-D86460608
2025-12-04T08:53:09.5160004Z  * [new branch]                export-D86474796        -> origin/export-D86474796
2025-12-04T08:53:09.5160180Z  * [new branch]                export-D86712396        -> origin/export-D86712396
2025-12-04T08:53:09.5160348Z  * [new branch]                export-D87022129        -> origin/export-D87022129
2025-12-04T08:53:09.5160559Z  * [new branch]                export-D87838959        -> origin/export-D87838959
2025-12-04T08:53:09.5160729Z  * [new branch]                export-D88319437        -> origin/export-D88319437
2025-12-04T08:53:09.5160944Z  * [new branch]                exported-model-train-idempotent -> origin/exported-model-train-idempotent
2025-12-04T08:53:09.5161176Z  * [new branch]                ezyang-titan-october    -> origin/ezyang-titan-october
2025-12-04T08:53:09.5161381Z  * [new branch]                ezyang-titan-october2   -> origin/ezyang-titan-october2
2025-12-04T08:53:09.5161564Z  * [new branch]                ezyang-war              -> origin/ezyang-war
2025-12-04T08:53:09.5161767Z  * [new branch]                ezyang/wip-aot-descriptors -> origin/ezyang/wip-aot-descriptors
2025-12-04T08:53:09.5161965Z  * [new branch]                fa_u8_brgemm            -> origin/fa_u8_brgemm
2025-12-04T08:53:09.5162214Z  * [new branch]                fadeputr/sequence_fbgemm -> origin/fadeputr/sequence_fbgemm
2025-12-04T08:53:09.5162408Z  * [new branch]                fastmath_baseline       -> origin/fastmath_baseline
2025-12-04T08:53:09.5162582Z  * [new branch]                fbcode/warm             -> origin/fbcode/warm
2025-12-04T08:53:09.5162796Z  * [new branch]                fca                     -> origin/fca
2025-12-04T08:53:09.5162955Z  * [new branch]                fca2_ca5984c            -> origin/fca2_ca5984c
2025-12-04T08:53:09.5163146Z  * [new branch]                fca5                    -> origin/fca5
2025-12-04T08:53:09.5163324Z  * [new branch]                feature/justknobs-cpp   -> origin/feature/justknobs-cpp
2025-12-04T08:53:09.5163533Z  * [new branch]                feature/numa-forkserver -> origin/feature/numa-forkserver
2025-12-04T08:53:09.5163723Z  * [new branch]                ffast_math_baseline     -> origin/ffast_math_baseline
2025-12-04T08:53:09.5163902Z  * [new branch]                ffast_math_target       -> origin/ffast_math_target
2025-12-04T08:53:09.5164082Z  * [new branch]                findhao/base_commit     -> origin/findhao/base_commit
2025-12-04T08:53:09.5164270Z  * [new branch]                findhao/base_commit1    -> origin/findhao/base_commit1
2025-12-04T08:53:09.5164457Z  * [new branch]                findhao/multistream2    -> origin/findhao/multistream2
2025-12-04T08:53:09.5164651Z  * [new branch]                findhao/multistream5    -> origin/findhao/multistream5
2025-12-04T08:53:09.5164837Z  * [new branch]                findhao/multistream6    -> origin/findhao/multistream6
2025-12-04T08:53:09.5165036Z  * [new branch]                findhao/operatorbench3  -> origin/findhao/operatorbench3
2025-12-04T08:53:09.5165237Z  * [new branch]                findhao/operatorbench5  -> origin/findhao/operatorbench5
2025-12-04T08:53:09.5165428Z  * [new branch]                findhao/tritonparse     -> origin/findhao/tritonparse
2025-12-04T08:53:09.5165640Z  * [new branch]                fix-ck-gemm-template-format -> origin/fix-ck-gemm-template-format
2025-12-04T08:53:09.5165854Z  * [new branch]                fix-config-ignore       -> origin/fix-config-ignore
2025-12-04T08:53:09.5166030Z  * [new branch]                fix-dict-guard          -> origin/fix-dict-guard
2025-12-04T08:53:09.5166204Z  * [new branch]                fix_addmm_issue         -> origin/fix_addmm_issue
2025-12-04T08:53:09.5166403Z  * [new branch]                fix_amd_missing_cluster_dims -> origin/fix_amd_missing_cluster_dims
2025-12-04T08:53:09.5166601Z  * [new branch]                fix_bench_bwd_pass      -> origin/fix_bench_bwd_pass
2025-12-04T08:53:09.5166793Z  * [new branch]                fix_mem_profiler_config -> origin/fix_mem_profiler_config
2025-12-04T08:53:09.5166977Z  * [new branch]                fix_nvrtc_discovery     -> origin/fix_nvrtc_discovery
2025-12-04T08:53:09.5167147Z  * [new branch]                fix_op_runner           -> origin/fix_op_runner
2025-12-04T08:53:09.5167314Z  * [new branch]                fix_ubn_159469          -> origin/fix_ubn_159469
2025-12-04T08:53:09.5167480Z  * [new branch]                fixes-triage            -> origin/fixes-triage
2025-12-04T08:53:09.5167656Z  * [new branch]                fixflashinfer           -> origin/fixflashinfer
2025-12-04T08:53:09.5167832Z  * [new branch]                flash_decoding_cpu      -> origin/flash_decoding_cpu
2025-12-04T08:53:09.5168006Z  * [new branch]                flex-flash              -> origin/flex-flash
2025-12-04T08:53:09.5168207Z  * [new branch]                flex_attention_functorch_grad -> origin/flex_attention_functorch_grad
2025-12-04T08:53:09.5168404Z  * [new branch]                flex_flash              -> origin/flex_flash
2025-12-04T08:53:09.5168607Z  * [new branch]                fmassa/fix_memeff_sharding_rule -> origin/fmassa/fix_memeff_sharding_rule
2025-12-04T08:53:09.5168851Z  * [new branch]                fmassa/tests_comm_compute_scheduler -> origin/fmassa/tests_comm_compute_scheduler
2025-12-04T08:53:09.5169067Z  * [new branch]                forkserver_fix          -> origin/forkserver_fix
2025-12-04T08:53:09.5169239Z  * [new branch]                fsdp2_trace_rules       -> origin/fsdp2_trace_rules
2025-12-04T08:53:09.5169418Z  * [new branch]                fx_cpp                  -> origin/fx_cpp
2025-12-04T08:53:09.5169611Z  * [new branch]                fy/fix-win              -> origin/fy/fix-win
2025-12-04T08:53:09.5169775Z  * [new branch]                galv-patch-1            -> origin/galv-patch-1
2025-12-04T08:53:09.5170049Z  * [new branch]                galv/cudagraphs-conditional-nodes-4 -> origin/galv/cudagraphs-conditional-nodes-4
2025-12-04T08:53:09.5170312Z  * [new branch]                georgehong/cmakelists-patch -> origin/georgehong/cmakelists-patch
2025-12-04T08:53:09.5170555Z  * [new branch]                gh/AlnisM/1/base        -> origin/gh/AlnisM/1/base
2025-12-04T08:53:09.5170733Z  * [new branch]                gh/AlnisM/1/head        -> origin/gh/AlnisM/1/head
2025-12-04T08:53:09.5170920Z  * [new branch]                gh/EikanWang/67/base    -> origin/gh/EikanWang/67/base
2025-12-04T08:53:09.5171120Z  * [new branch]                gh/EikanWang/67/head    -> origin/gh/EikanWang/67/head
2025-12-04T08:53:09.5171316Z  * [new branch]                gh/Gasoonjia/1/base     -> origin/gh/Gasoonjia/1/base
2025-12-04T08:53:09.5171508Z  * [new branch]                gh/Gasoonjia/1/head     -> origin/gh/Gasoonjia/1/head
2025-12-04T08:53:09.5171694Z  * [new branch]                gh/H-Huang/131/base     -> origin/gh/H-Huang/131/base
2025-12-04T08:53:09.5171885Z  * [new branch]                gh/H-Huang/131/head     -> origin/gh/H-Huang/131/head
2025-12-04T08:53:09.5172069Z  * [new branch]                gh/H-Huang/131/orig     -> origin/gh/H-Huang/131/orig
2025-12-04T08:53:09.5172252Z  * [new branch]                gh/H-Huang/132/base     -> origin/gh/H-Huang/132/base
2025-12-04T08:53:09.5172435Z  * [new branch]                gh/H-Huang/132/head     -> origin/gh/H-Huang/132/head
2025-12-04T08:53:09.5172614Z  * [new branch]                gh/H-Huang/132/orig     -> origin/gh/H-Huang/132/orig
2025-12-04T08:53:09.5172804Z  * [new branch]                gh/H-Huang/180/base     -> origin/gh/H-Huang/180/base
2025-12-04T08:53:09.5172995Z  * [new branch]                gh/H-Huang/180/head     -> origin/gh/H-Huang/180/head
2025-12-04T08:53:09.5173184Z  * [new branch]                gh/H-Huang/180/orig     -> origin/gh/H-Huang/180/orig
2025-12-04T08:53:09.5173368Z  * [new branch]                gh/H-Huang/182/base     -> origin/gh/H-Huang/182/base
2025-12-04T08:53:09.5173555Z  * [new branch]                gh/H-Huang/182/head     -> origin/gh/H-Huang/182/head
2025-12-04T08:53:09.5173733Z  * [new branch]                gh/H-Huang/182/orig     -> origin/gh/H-Huang/182/orig
2025-12-04T08:53:09.5173925Z  * [new branch]                gh/H-Huang/226/base     -> origin/gh/H-Huang/226/base
2025-12-04T08:53:09.5174109Z  * [new branch]                gh/H-Huang/226/head     -> origin/gh/H-Huang/226/head
2025-12-04T08:53:09.5174287Z  * [new branch]                gh/H-Huang/226/orig     -> origin/gh/H-Huang/226/orig
2025-12-04T08:53:09.5174467Z  * [new branch]                gh/H-Huang/228/base     -> origin/gh/H-Huang/228/base
2025-12-04T08:53:09.5174654Z  * [new branch]                gh/H-Huang/228/head     -> origin/gh/H-Huang/228/head
2025-12-04T08:53:09.5174835Z  * [new branch]                gh/H-Huang/228/orig     -> origin/gh/H-Huang/228/orig
2025-12-04T08:53:09.5175036Z  * [new branch]                gh/IvanKobzarev/150/base -> origin/gh/IvanKobzarev/150/base
2025-12-04T08:53:09.5175247Z  * [new branch]                gh/IvanKobzarev/150/head -> origin/gh/IvanKobzarev/150/head
2025-12-04T08:53:09.5175455Z  * [new branch]                gh/IvanKobzarev/150/orig -> origin/gh/IvanKobzarev/150/orig
2025-12-04T08:53:09.5175660Z  * [new branch]                gh/IvanKobzarev/157/base -> origin/gh/IvanKobzarev/157/base
2025-12-04T08:53:09.5175862Z  * [new branch]                gh/IvanKobzarev/157/head -> origin/gh/IvanKobzarev/157/head
2025-12-04T08:53:09.5176064Z  * [new branch]                gh/IvanKobzarev/157/orig -> origin/gh/IvanKobzarev/157/orig
2025-12-04T08:53:09.5176267Z  * [new branch]                gh/IvanKobzarev/159/base -> origin/gh/IvanKobzarev/159/base
2025-12-04T08:53:09.5176466Z  * [new branch]                gh/IvanKobzarev/159/head -> origin/gh/IvanKobzarev/159/head
2025-12-04T08:53:09.5176707Z  * [new branch]                gh/IvanKobzarev/159/orig -> origin/gh/IvanKobzarev/159/orig
2025-12-04T08:53:09.5176939Z  * [new branch]                gh/IvanKobzarev/162/base -> origin/gh/IvanKobzarev/162/base
2025-12-04T08:53:09.5177138Z  * [new branch]                gh/IvanKobzarev/162/head -> origin/gh/IvanKobzarev/162/head
2025-12-04T08:53:09.5177340Z  * [new branch]                gh/IvanKobzarev/162/orig -> origin/gh/IvanKobzarev/162/orig
2025-12-04T08:53:09.5177547Z  * [new branch]                gh/IvanKobzarev/163/base -> origin/gh/IvanKobzarev/163/base
2025-12-04T08:53:09.5177748Z  * [new branch]                gh/IvanKobzarev/163/head -> origin/gh/IvanKobzarev/163/head
2025-12-04T08:53:09.5177958Z  * [new branch]                gh/IvanKobzarev/163/orig -> origin/gh/IvanKobzarev/163/orig
2025-12-04T08:53:09.5178167Z  * [new branch]                gh/IvanKobzarev/166/base -> origin/gh/IvanKobzarev/166/base
2025-12-04T08:53:09.5178373Z  * [new branch]                gh/IvanKobzarev/166/head -> origin/gh/IvanKobzarev/166/head
2025-12-04T08:53:09.5178580Z  * [new branch]                gh/IvanKobzarev/166/orig -> origin/gh/IvanKobzarev/166/orig
2025-12-04T08:53:09.5178784Z  * [new branch]                gh/IvanKobzarev/167/base -> origin/gh/IvanKobzarev/167/base
2025-12-04T08:53:09.5178983Z  * [new branch]                gh/IvanKobzarev/167/head -> origin/gh/IvanKobzarev/167/head
2025-12-04T08:53:09.5179192Z  * [new branch]                gh/IvanKobzarev/167/orig -> origin/gh/IvanKobzarev/167/orig
2025-12-04T08:53:09.5179395Z  * [new branch]                gh/IvanKobzarev/168/base -> origin/gh/IvanKobzarev/168/base
2025-12-04T08:53:09.5179593Z  * [new branch]                gh/IvanKobzarev/168/head -> origin/gh/IvanKobzarev/168/head
2025-12-04T08:53:09.5179796Z  * [new branch]                gh/IvanKobzarev/168/orig -> origin/gh/IvanKobzarev/168/orig
2025-12-04T08:53:09.5180004Z  * [new branch]                gh/IvanKobzarev/169/base -> origin/gh/IvanKobzarev/169/base
2025-12-04T08:53:09.5180205Z  * [new branch]                gh/IvanKobzarev/169/head -> origin/gh/IvanKobzarev/169/head
2025-12-04T08:53:09.5180454Z  * [new branch]                gh/IvanKobzarev/169/orig -> origin/gh/IvanKobzarev/169/orig
2025-12-04T08:53:09.5180659Z  * [new branch]                gh/IvanKobzarev/170/base -> origin/gh/IvanKobzarev/170/base
2025-12-04T08:53:09.5180859Z  * [new branch]                gh/IvanKobzarev/170/head -> origin/gh/IvanKobzarev/170/head
2025-12-04T08:53:09.5181067Z  * [new branch]                gh/IvanKobzarev/170/orig -> origin/gh/IvanKobzarev/170/orig
2025-12-04T08:53:09.5181279Z  * [new branch]                gh/IvanKobzarev/171/base -> origin/gh/IvanKobzarev/171/base
2025-12-04T08:53:09.5181478Z  * [new branch]                gh/IvanKobzarev/171/head -> origin/gh/IvanKobzarev/171/head
2025-12-04T08:53:09.5181681Z  * [new branch]                gh/IvanKobzarev/171/orig -> origin/gh/IvanKobzarev/171/orig
2025-12-04T08:53:09.5181895Z  * [new branch]                gh/IvanKobzarev/172/base -> origin/gh/IvanKobzarev/172/base
2025-12-04T08:53:09.5182095Z  * [new branch]                gh/IvanKobzarev/172/head -> origin/gh/IvanKobzarev/172/head
2025-12-04T08:53:09.5182303Z  * [new branch]                gh/IvanKobzarev/172/orig -> origin/gh/IvanKobzarev/172/orig
2025-12-04T08:53:09.5182502Z  * [new branch]                gh/IvanKobzarev/173/base -> origin/gh/IvanKobzarev/173/base
2025-12-04T08:53:09.5182709Z  * [new branch]                gh/IvanKobzarev/173/head -> origin/gh/IvanKobzarev/173/head
2025-12-04T08:53:09.5182911Z  * [new branch]                gh/IvanKobzarev/173/orig -> origin/gh/IvanKobzarev/173/orig
2025-12-04T08:53:09.5183110Z  * [new branch]                gh/IvanKobzarev/174/base -> origin/gh/IvanKobzarev/174/base
2025-12-04T08:53:09.5183313Z  * [new branch]                gh/IvanKobzarev/174/head -> origin/gh/IvanKobzarev/174/head
2025-12-04T08:53:09.5183516Z  * [new branch]                gh/IvanKobzarev/174/orig -> origin/gh/IvanKobzarev/174/orig
2025-12-04T08:53:09.5183772Z  * [new branch]                gh/IvanKobzarev/175/base -> origin/gh/IvanKobzarev/175/base
2025-12-04T08:53:09.5184004Z  * [new branch]                gh/IvanKobzarev/175/head -> origin/gh/IvanKobzarev/175/head
2025-12-04T08:53:09.5184205Z  * [new branch]                gh/IvanKobzarev/175/orig -> origin/gh/IvanKobzarev/175/orig
2025-12-04T08:53:09.5184401Z  * [new branch]                gh/IvanKobzarev/176/base -> origin/gh/IvanKobzarev/176/base
2025-12-04T08:53:09.5184611Z  * [new branch]                gh/IvanKobzarev/176/head -> origin/gh/IvanKobzarev/176/head
2025-12-04T08:53:09.5184815Z  * [new branch]                gh/IvanKobzarev/176/orig -> origin/gh/IvanKobzarev/176/orig
2025-12-04T08:53:09.5185009Z  * [new branch]                gh/IvanKobzarev/177/base -> origin/gh/IvanKobzarev/177/base
2025-12-04T08:53:09.5185206Z  * [new branch]                gh/IvanKobzarev/177/head -> origin/gh/IvanKobzarev/177/head
2025-12-04T08:53:09.5185416Z  * [new branch]                gh/IvanKobzarev/177/orig -> origin/gh/IvanKobzarev/177/orig
2025-12-04T08:53:09.5185613Z  * [new branch]                gh/IvanKobzarev/178/base -> origin/gh/IvanKobzarev/178/base
2025-12-04T08:53:09.5185813Z  * [new branch]                gh/IvanKobzarev/178/head -> origin/gh/IvanKobzarev/178/head
2025-12-04T08:53:09.5186010Z  * [new branch]                gh/IvanKobzarev/178/orig -> origin/gh/IvanKobzarev/178/orig
2025-12-04T08:53:09.5186212Z  * [new branch]                gh/IvanKobzarev/179/base -> origin/gh/IvanKobzarev/179/base
2025-12-04T08:53:09.5186417Z  * [new branch]                gh/IvanKobzarev/179/head -> origin/gh/IvanKobzarev/179/head
2025-12-04T08:53:09.5186614Z  * [new branch]                gh/IvanKobzarev/179/orig -> origin/gh/IvanKobzarev/179/orig
2025-12-04T08:53:09.5186808Z  * [new branch]                gh/IvanKobzarev/180/base -> origin/gh/IvanKobzarev/180/base
2025-12-04T08:53:09.5187008Z  * [new branch]                gh/IvanKobzarev/180/head -> origin/gh/IvanKobzarev/180/head
2025-12-04T08:53:09.5187210Z  * [new branch]                gh/IvanKobzarev/180/orig -> origin/gh/IvanKobzarev/180/orig
2025-12-04T08:53:09.5187415Z  * [new branch]                gh/IvanKobzarev/181/base -> origin/gh/IvanKobzarev/181/base
2025-12-04T08:53:09.5187613Z  * [new branch]                gh/IvanKobzarev/181/head -> origin/gh/IvanKobzarev/181/head
2025-12-04T08:53:09.5187810Z  * [new branch]                gh/IvanKobzarev/181/orig -> origin/gh/IvanKobzarev/181/orig
2025-12-04T08:53:09.5188009Z  * [new branch]                gh/IvanKobzarev/182/base -> origin/gh/IvanKobzarev/182/base
2025-12-04T08:53:09.5188206Z  * [new branch]                gh/IvanKobzarev/182/head -> origin/gh/IvanKobzarev/182/head
2025-12-04T08:53:09.5188406Z  * [new branch]                gh/IvanKobzarev/182/orig -> origin/gh/IvanKobzarev/182/orig
2025-12-04T08:53:09.5188605Z  * [new branch]                gh/IvanKobzarev/183/base -> origin/gh/IvanKobzarev/183/base
2025-12-04T08:53:09.5188812Z  * [new branch]                gh/IvanKobzarev/183/head -> origin/gh/IvanKobzarev/183/head
2025-12-04T08:53:09.5189020Z  * [new branch]                gh/IvanKobzarev/183/orig -> origin/gh/IvanKobzarev/183/orig
2025-12-04T08:53:09.5189221Z  * [new branch]                gh/IvanKobzarev/184/base -> origin/gh/IvanKobzarev/184/base
2025-12-04T08:53:09.5189427Z  * [new branch]                gh/IvanKobzarev/184/head -> origin/gh/IvanKobzarev/184/head
2025-12-04T08:53:09.5189626Z  * [new branch]                gh/IvanKobzarev/184/orig -> origin/gh/IvanKobzarev/184/orig
2025-12-04T08:53:09.5189835Z  * [new branch]                gh/NikhilAPatel/1/base  -> origin/gh/NikhilAPatel/1/base
2025-12-04T08:53:09.5190040Z  * [new branch]                gh/NikhilAPatel/1/head  -> origin/gh/NikhilAPatel/1/head
2025-12-04T08:53:09.5190239Z  * [new branch]                gh/NikhilAPatel/2/base  -> origin/gh/NikhilAPatel/2/base
2025-12-04T08:53:09.5190472Z  * [new branch]                gh/NikhilAPatel/2/head  -> origin/gh/NikhilAPatel/2/head
2025-12-04T08:53:09.5190716Z  * [new branch]                gh/NikhilAPatel/4/base  -> origin/gh/NikhilAPatel/4/base
2025-12-04T08:53:09.5190951Z  * [new branch]                gh/NikhilAPatel/4/head  -> origin/gh/NikhilAPatel/4/head
2025-12-04T08:53:09.5191150Z  * [new branch]                gh/NikhilAPatel/5/base  -> origin/gh/NikhilAPatel/5/base
2025-12-04T08:53:09.5191348Z  * [new branch]                gh/NikhilAPatel/5/head  -> origin/gh/NikhilAPatel/5/head
2025-12-04T08:53:09.5191547Z  * [new branch]                gh/NikhilAPatel/5/orig  -> origin/gh/NikhilAPatel/5/orig
2025-12-04T08:53:09.5191743Z  * [new branch]                gh/PaliC/17/base        -> origin/gh/PaliC/17/base
2025-12-04T08:53:09.5191927Z  * [new branch]                gh/PaliC/17/head        -> origin/gh/PaliC/17/head
2025-12-04T08:53:09.5192107Z  * [new branch]                gh/PaliC/17/orig        -> origin/gh/PaliC/17/orig
2025-12-04T08:53:09.5192294Z  * [new branch]                gh/PaliC/18/base        -> origin/gh/PaliC/18/base
2025-12-04T08:53:09.5192481Z  * [new branch]                gh/PaliC/18/head        -> origin/gh/PaliC/18/head
2025-12-04T08:53:09.5192656Z  * [new branch]                gh/PaliC/18/orig        -> origin/gh/PaliC/18/orig
2025-12-04T08:53:09.5192836Z  * [new branch]                gh/PaliC/20/base        -> origin/gh/PaliC/20/base
2025-12-04T08:53:09.5193014Z  * [new branch]                gh/PaliC/20/head        -> origin/gh/PaliC/20/head
2025-12-04T08:53:09.5193190Z  * [new branch]                gh/PaliC/20/orig        -> origin/gh/PaliC/20/orig
2025-12-04T08:53:09.5193369Z  * [new branch]                gh/PaliC/21/base        -> origin/gh/PaliC/21/base
2025-12-04T08:53:09.5193545Z  * [new branch]                gh/PaliC/21/head        -> origin/gh/PaliC/21/head
2025-12-04T08:53:09.5193718Z  * [new branch]                gh/PaliC/21/orig        -> origin/gh/PaliC/21/orig
2025-12-04T08:53:09.5193900Z  * [new branch]                gh/PaliC/23/base        -> origin/gh/PaliC/23/base
2025-12-04T08:53:09.5194078Z  * [new branch]                gh/PaliC/23/head        -> origin/gh/PaliC/23/head
2025-12-04T08:53:09.5194255Z  * [new branch]                gh/PaliC/23/orig        -> origin/gh/PaliC/23/orig
2025-12-04T08:53:09.5194434Z  * [new branch]                gh/PaliC/24/base        -> origin/gh/PaliC/24/base
2025-12-04T08:53:09.5194607Z  * [new branch]                gh/PaliC/24/head        -> origin/gh/PaliC/24/head
2025-12-04T08:53:09.5194789Z  * [new branch]                gh/PaliC/24/orig        -> origin/gh/PaliC/24/orig
2025-12-04T08:53:09.5194967Z  * [new branch]                gh/PaliC/25/head        -> origin/gh/PaliC/25/head
2025-12-04T08:53:09.5195140Z  * [new branch]                gh/PaliC/25/next        -> origin/gh/PaliC/25/next
2025-12-04T08:53:09.5195316Z  * [new branch]                gh/PaliC/25/orig        -> origin/gh/PaliC/25/orig
2025-12-04T08:53:09.5195492Z  * [new branch]                gh/PaliC/26/head        -> origin/gh/PaliC/26/head
2025-12-04T08:53:09.5195671Z  * [new branch]                gh/PaliC/26/next        -> origin/gh/PaliC/26/next
2025-12-04T08:53:09.5195847Z  * [new branch]                gh/PaliC/26/orig        -> origin/gh/PaliC/26/orig
2025-12-04T08:53:09.5196027Z  * [new branch]                gh/PaliC/27/next        -> origin/gh/PaliC/27/next
2025-12-04T08:53:09.5196199Z  * [new branch]                gh/PaliC/28/head        -> origin/gh/PaliC/28/head
2025-12-04T08:53:09.5196373Z  * [new branch]                gh/PaliC/28/next        -> origin/gh/PaliC/28/next
2025-12-04T08:53:09.5196557Z  * [new branch]                gh/PaliC/28/orig        -> origin/gh/PaliC/28/orig
2025-12-04T08:53:09.5196731Z  * [new branch]                gh/PaliC/29/head        -> origin/gh/PaliC/29/head
2025-12-04T08:53:09.5196909Z  * [new branch]                gh/PaliC/29/next        -> origin/gh/PaliC/29/next
2025-12-04T08:53:09.5197087Z  * [new branch]                gh/PaliC/29/orig        -> origin/gh/PaliC/29/orig
2025-12-04T08:53:09.5197260Z  * [new branch]                gh/PaliC/30/head        -> origin/gh/PaliC/30/head
2025-12-04T08:53:09.5197475Z  * [new branch]                gh/PaliC/30/next        -> origin/gh/PaliC/30/next
2025-12-04T08:53:09.5197682Z  * [new branch]                gh/PaliC/30/orig        -> origin/gh/PaliC/30/orig
2025-12-04T08:53:09.5197860Z  * [new branch]                gh/PaliC/31/head        -> origin/gh/PaliC/31/head
2025-12-04T08:53:09.5198036Z  * [new branch]                gh/PaliC/31/next        -> origin/gh/PaliC/31/next
2025-12-04T08:53:09.5198216Z  * [new branch]                gh/PaliC/31/orig        -> origin/gh/PaliC/31/orig
2025-12-04T08:53:09.5198407Z  * [new branch]                gh/PaulZhang12/25/base  -> origin/gh/PaulZhang12/25/base
2025-12-04T08:53:09.5198608Z  * [new branch]                gh/PaulZhang12/25/head  -> origin/gh/PaulZhang12/25/head
2025-12-04T08:53:09.5198800Z  * [new branch]                gh/PaulZhang12/25/orig  -> origin/gh/PaulZhang12/25/orig
2025-12-04T08:53:09.5199001Z  * [new branch]                gh/PaulZhang12/28/base  -> origin/gh/PaulZhang12/28/base
2025-12-04T08:53:09.5199202Z  * [new branch]                gh/PaulZhang12/28/head  -> origin/gh/PaulZhang12/28/head
2025-12-04T08:53:09.5199396Z  * [new branch]                gh/PaulZhang12/28/orig  -> origin/gh/PaulZhang12/28/orig
2025-12-04T08:53:09.5199593Z  * [new branch]                gh/PaulZhang12/31/base  -> origin/gh/PaulZhang12/31/base
2025-12-04T08:53:09.5199789Z  * [new branch]                gh/PaulZhang12/31/head  -> origin/gh/PaulZhang12/31/head
2025-12-04T08:53:09.5199983Z  * [new branch]                gh/PaulZhang12/31/orig  -> origin/gh/PaulZhang12/31/orig
2025-12-04T08:53:09.5200180Z  * [new branch]                gh/PaulZhang12/37/base  -> origin/gh/PaulZhang12/37/base
2025-12-04T08:53:09.5200376Z  * [new branch]                gh/PaulZhang12/37/head  -> origin/gh/PaulZhang12/37/head
2025-12-04T08:53:09.5200617Z  * [new branch]                gh/PaulZhang12/37/orig  -> origin/gh/PaulZhang12/37/orig
2025-12-04T08:53:09.5200818Z  * [new branch]                gh/PaulZhang12/40/base  -> origin/gh/PaulZhang12/40/base
2025-12-04T08:53:09.5201017Z  * [new branch]                gh/PaulZhang12/40/head  -> origin/gh/PaulZhang12/40/head
2025-12-04T08:53:09.5201211Z  * [new branch]                gh/PaulZhang12/40/orig  -> origin/gh/PaulZhang12/40/orig
2025-12-04T08:53:09.5201407Z  * [new branch]                gh/PaulZhang12/42/base  -> origin/gh/PaulZhang12/42/base
2025-12-04T08:53:09.5201608Z  * [new branch]                gh/PaulZhang12/42/head  -> origin/gh/PaulZhang12/42/head
2025-12-04T08:53:09.5201799Z  * [new branch]                gh/PaulZhang12/43/base  -> origin/gh/PaulZhang12/43/base
2025-12-04T08:53:09.5201999Z  * [new branch]                gh/PaulZhang12/43/head  -> origin/gh/PaulZhang12/43/head
2025-12-04T08:53:09.5202191Z  * [new branch]                gh/PaulZhang12/43/orig  -> origin/gh/PaulZhang12/43/orig
2025-12-04T08:53:09.5202393Z  * [new branch]                gh/PaulZhang12/44/base  -> origin/gh/PaulZhang12/44/base
2025-12-04T08:53:09.5202593Z  * [new branch]                gh/PaulZhang12/44/head  -> origin/gh/PaulZhang12/44/head
2025-12-04T08:53:09.5202786Z  * [new branch]                gh/PaulZhang12/45/base  -> origin/gh/PaulZhang12/45/base
2025-12-04T08:53:09.5202982Z  * [new branch]                gh/PaulZhang12/45/head  -> origin/gh/PaulZhang12/45/head
2025-12-04T08:53:09.5203183Z  * [new branch]                gh/PaulZhang12/45/orig  -> origin/gh/PaulZhang12/45/orig
2025-12-04T08:53:09.5203377Z  * [new branch]                gh/PaulZhang12/46/base  -> origin/gh/PaulZhang12/46/base
2025-12-04T08:53:09.5203571Z  * [new branch]                gh/PaulZhang12/46/head  -> origin/gh/PaulZhang12/46/head
2025-12-04T08:53:09.5203765Z  * [new branch]                gh/PaulZhang12/46/orig  -> origin/gh/PaulZhang12/46/orig
2025-12-04T08:53:09.5203959Z  * [new branch]                gh/PaulZhang12/47/base  -> origin/gh/PaulZhang12/47/base
2025-12-04T08:53:09.5204152Z  * [new branch]                gh/PaulZhang12/47/head  -> origin/gh/PaulZhang12/47/head
2025-12-04T08:53:09.5204394Z  * [new branch]                gh/PaulZhang12/47/orig  -> origin/gh/PaulZhang12/47/orig
2025-12-04T08:53:09.5204616Z  * [new branch]                gh/PaulZhang12/48/base  -> origin/gh/PaulZhang12/48/base
2025-12-04T08:53:09.5204814Z  * [new branch]                gh/PaulZhang12/48/head  -> origin/gh/PaulZhang12/48/head
2025-12-04T08:53:09.5205008Z  * [new branch]                gh/PaulZhang12/48/orig  -> origin/gh/PaulZhang12/48/orig
2025-12-04T08:53:09.5205201Z  * [new branch]                gh/SamGinzburg/11/base  -> origin/gh/SamGinzburg/11/base
2025-12-04T08:53:09.5205397Z  * [new branch]                gh/SamGinzburg/11/head  -> origin/gh/SamGinzburg/11/head
2025-12-04T08:53:09.5205607Z  * [new branch]                gh/SherlockNoMad/1/base -> origin/gh/SherlockNoMad/1/base
2025-12-04T08:53:09.5205806Z  * [new branch]                gh/SherlockNoMad/1/head -> origin/gh/SherlockNoMad/1/head
2025-12-04T08:53:09.5206013Z  * [new branch]                gh/SherlockNoMad/10/base -> origin/gh/SherlockNoMad/10/base
2025-12-04T08:53:09.5206228Z  * [new branch]                gh/SherlockNoMad/10/head -> origin/gh/SherlockNoMad/10/head
2025-12-04T08:53:09.5206438Z  * [new branch]                gh/SherlockNoMad/10/orig -> origin/gh/SherlockNoMad/10/orig
2025-12-04T08:53:09.5206645Z  * [new branch]                gh/SherlockNoMad/11/base -> origin/gh/SherlockNoMad/11/base
2025-12-04T08:53:09.5206850Z  * [new branch]                gh/SherlockNoMad/11/head -> origin/gh/SherlockNoMad/11/head
2025-12-04T08:53:09.5207050Z  * [new branch]                gh/SherlockNoMad/11/orig -> origin/gh/SherlockNoMad/11/orig
2025-12-04T08:53:09.5207258Z  * [new branch]                gh/SherlockNoMad/12/base -> origin/gh/SherlockNoMad/12/base
2025-12-04T08:53:09.5207463Z  * [new branch]                gh/SherlockNoMad/12/head -> origin/gh/SherlockNoMad/12/head
2025-12-04T08:53:09.5207662Z  * [new branch]                gh/SherlockNoMad/12/orig -> origin/gh/SherlockNoMad/12/orig
2025-12-04T08:53:09.5207867Z  * [new branch]                gh/SherlockNoMad/15/base -> origin/gh/SherlockNoMad/15/base
2025-12-04T08:53:09.5208068Z  * [new branch]                gh/SherlockNoMad/15/head -> origin/gh/SherlockNoMad/15/head
2025-12-04T08:53:09.5208279Z  * [new branch]                gh/SherlockNoMad/15/orig -> origin/gh/SherlockNoMad/15/orig
2025-12-04T08:53:09.5208489Z  * [new branch]                gh/SherlockNoMad/17/base -> origin/gh/SherlockNoMad/17/base
2025-12-04T08:53:09.5208689Z  * [new branch]                gh/SherlockNoMad/17/head -> origin/gh/SherlockNoMad/17/head
2025-12-04T08:53:09.5208899Z  * [new branch]                gh/SherlockNoMad/17/orig -> origin/gh/SherlockNoMad/17/orig
2025-12-04T08:53:09.5209105Z  * [new branch]                gh/SherlockNoMad/18/base -> origin/gh/SherlockNoMad/18/base
2025-12-04T08:53:09.5209306Z  * [new branch]                gh/SherlockNoMad/18/head -> origin/gh/SherlockNoMad/18/head
2025-12-04T08:53:09.5209511Z  * [new branch]                gh/SherlockNoMad/18/orig -> origin/gh/SherlockNoMad/18/orig
2025-12-04T08:53:09.5209731Z  * [new branch]                gh/SherlockNoMad/19/base -> origin/gh/SherlockNoMad/19/base
2025-12-04T08:53:09.5209932Z  * [new branch]                gh/SherlockNoMad/19/head -> origin/gh/SherlockNoMad/19/head
2025-12-04T08:53:09.5210139Z  * [new branch]                gh/SherlockNoMad/19/orig -> origin/gh/SherlockNoMad/19/orig
2025-12-04T08:53:09.5210345Z  * [new branch]                gh/SherlockNoMad/2/base -> origin/gh/SherlockNoMad/2/base
2025-12-04T08:53:09.5210588Z  * [new branch]                gh/SherlockNoMad/2/head -> origin/gh/SherlockNoMad/2/head
2025-12-04T08:53:09.5210791Z  * [new branch]                gh/SherlockNoMad/20/base -> origin/gh/SherlockNoMad/20/base
2025-12-04T08:53:09.5210997Z  * [new branch]                gh/SherlockNoMad/20/head -> origin/gh/SherlockNoMad/20/head
2025-12-04T08:53:09.5211197Z  * [new branch]                gh/SherlockNoMad/20/orig -> origin/gh/SherlockNoMad/20/orig
2025-12-04T08:53:09.5211443Z  * [new branch]                gh/SherlockNoMad/21/base -> origin/gh/SherlockNoMad/21/base
2025-12-04T08:53:09.5211647Z  * [new branch]                gh/SherlockNoMad/21/head -> origin/gh/SherlockNoMad/21/head
2025-12-04T08:53:09.5211884Z  * [new branch]                gh/SherlockNoMad/21/orig -> origin/gh/SherlockNoMad/21/orig
2025-12-04T08:53:09.5212091Z  * [new branch]                gh/SherlockNoMad/3/base -> origin/gh/SherlockNoMad/3/base
2025-12-04T08:53:09.5212295Z  * [new branch]                gh/SherlockNoMad/3/head -> origin/gh/SherlockNoMad/3/head
2025-12-04T08:53:09.5212491Z  * [new branch]                gh/SherlockNoMad/4/base -> origin/gh/SherlockNoMad/4/base
2025-12-04T08:53:09.5212689Z  * [new branch]                gh/SherlockNoMad/4/head -> origin/gh/SherlockNoMad/4/head
2025-12-04T08:53:09.5213043Z  * [new branch]                gh/SherlockNoMad/5/base -> origin/gh/SherlockNoMad/5/base
2025-12-04T08:53:09.5213571Z  * [new branch]                gh/SherlockNoMad/5/head -> origin/gh/SherlockNoMad/5/head
2025-12-04T08:53:09.5213835Z  * [new branch]                gh/Sidharth123-cpu/24/base -> origin/gh/Sidharth123-cpu/24/base
2025-12-04T08:53:09.5214079Z  * [new branch]                gh/Sidharth123-cpu/25/base -> origin/gh/Sidharth123-cpu/25/base
2025-12-04T08:53:09.5214363Z  * [new branch]                gh/Sidharth123-cpu/26/base -> origin/gh/Sidharth123-cpu/26/base
2025-12-04T08:53:09.5233637Z  * [new branch]                gh/Sidharth123-cpu/27/base -> origin/gh/Sidharth123-cpu/27/base
2025-12-04T08:53:09.5233914Z  * [new branch]                gh/StrongerXi/1/base    -> origin/gh/StrongerXi/1/base
2025-12-04T08:53:09.5234119Z  * [new branch]                gh/StrongerXi/1/head    -> origin/gh/StrongerXi/1/head
2025-12-04T08:53:09.5234342Z  * [new branch]                gh/StrongerXi/71/base   -> origin/gh/StrongerXi/71/base
2025-12-04T08:53:09.5234555Z  * [new branch]                gh/StrongerXi/71/head   -> origin/gh/StrongerXi/71/head
2025-12-04T08:53:09.5234754Z  * [new branch]                gh/StrongerXi/72/base   -> origin/gh/StrongerXi/72/base
2025-12-04T08:53:09.5234994Z  * [new branch]                gh/StrongerXi/72/head   -> origin/gh/StrongerXi/72/head
2025-12-04T08:53:09.5235212Z  * [new branch]                gh/StrongerXi/73/base   -> origin/gh/StrongerXi/73/base
2025-12-04T08:53:09.5235420Z  * [new branch]                gh/StrongerXi/73/head   -> origin/gh/StrongerXi/73/head
2025-12-04T08:53:09.5235617Z  * [new branch]                gh/StrongerXi/73/orig   -> origin/gh/StrongerXi/73/orig
2025-12-04T08:53:09.5235805Z  * [new branch]                gh/XilunWu/160/base     -> origin/gh/XilunWu/160/base
2025-12-04T08:53:09.5236001Z  * [new branch]                gh/XilunWu/160/head     -> origin/gh/XilunWu/160/head
2025-12-04T08:53:09.5236192Z  * [new branch]                gh/XilunWu/160/orig     -> origin/gh/XilunWu/160/orig
2025-12-04T08:53:09.5236372Z  * [new branch]                gh/XilunWu/163/base     -> origin/gh/XilunWu/163/base
2025-12-04T08:53:09.5236564Z  * [new branch]                gh/XilunWu/163/head     -> origin/gh/XilunWu/163/head
2025-12-04T08:53:09.5236750Z  * [new branch]                gh/XilunWu/163/orig     -> origin/gh/XilunWu/163/orig
2025-12-04T08:53:09.5236936Z  * [new branch]                gh/XilunWu/168/base     -> origin/gh/XilunWu/168/base
2025-12-04T08:53:09.5237127Z  * [new branch]                gh/XilunWu/168/head     -> origin/gh/XilunWu/168/head
2025-12-04T08:53:09.5237314Z  * [new branch]                gh/XilunWu/168/orig     -> origin/gh/XilunWu/168/orig
2025-12-04T08:53:09.5237499Z  * [new branch]                gh/XilunWu/169/base     -> origin/gh/XilunWu/169/base
2025-12-04T08:53:09.5237691Z  * [new branch]                gh/XilunWu/169/head     -> origin/gh/XilunWu/169/head
2025-12-04T08:53:09.5237872Z  * [new branch]                gh/XilunWu/169/orig     -> origin/gh/XilunWu/169/orig
2025-12-04T08:53:09.5238073Z  * [new branch]                gh/XilunWu/170/base     -> origin/gh/XilunWu/170/base
2025-12-04T08:53:09.5238348Z  * [new branch]                gh/XilunWu/170/head     -> origin/gh/XilunWu/170/head
2025-12-04T08:53:09.5238538Z  * [new branch]                gh/XilunWu/170/orig     -> origin/gh/XilunWu/170/orig
2025-12-04T08:53:09.5238754Z  * [new branch]                gh/XilunWu/171/base     -> origin/gh/XilunWu/171/base
2025-12-04T08:53:09.5238944Z  * [new branch]                gh/XilunWu/171/head     -> origin/gh/XilunWu/171/head
2025-12-04T08:53:09.5239130Z  * [new branch]                gh/XilunWu/171/orig     -> origin/gh/XilunWu/171/orig
2025-12-04T08:53:09.5239315Z  * [new branch]                gh/XilunWu/173/base     -> origin/gh/XilunWu/173/base
2025-12-04T08:53:09.5239510Z  * [new branch]                gh/XilunWu/173/head     -> origin/gh/XilunWu/173/head
2025-12-04T08:53:09.5239686Z  * [new branch]                gh/XilunWu/173/orig     -> origin/gh/XilunWu/173/orig
2025-12-04T08:53:09.5239881Z  * [new branch]                gh/XilunWu/175/base     -> origin/gh/XilunWu/175/base
2025-12-04T08:53:09.5240068Z  * [new branch]                gh/XilunWu/175/head     -> origin/gh/XilunWu/175/head
2025-12-04T08:53:09.5240255Z  * [new branch]                gh/XilunWu/175/orig     -> origin/gh/XilunWu/175/orig
2025-12-04T08:53:09.5240510Z  * [new branch]                gh/XilunWu/176/base     -> origin/gh/XilunWu/176/base
2025-12-04T08:53:09.5240692Z  * [new branch]                gh/XilunWu/176/head     -> origin/gh/XilunWu/176/head
2025-12-04T08:53:09.5240867Z  * [new branch]                gh/XilunWu/176/orig     -> origin/gh/XilunWu/176/orig
2025-12-04T08:53:09.5241063Z  * [new branch]                gh/XuehaiPan/14/base    -> origin/gh/XuehaiPan/14/base
2025-12-04T08:53:09.5241248Z  * [new branch]                gh/XuehaiPan/14/head    -> origin/gh/XuehaiPan/14/head
2025-12-04T08:53:09.5241433Z  * [new branch]                gh/XuehaiPan/14/orig    -> origin/gh/XuehaiPan/14/orig
2025-12-04T08:53:09.5241618Z  * [new branch]                gh/XuehaiPan/179/base   -> origin/gh/XuehaiPan/179/base
2025-12-04T08:53:09.5241808Z  * [new branch]                gh/XuehaiPan/179/head   -> origin/gh/XuehaiPan/179/head
2025-12-04T08:53:09.5241995Z  * [new branch]                gh/XuehaiPan/179/orig   -> origin/gh/XuehaiPan/179/orig
2025-12-04T08:53:09.5242187Z  * [new branch]                gh/XuehaiPan/249/base   -> origin/gh/XuehaiPan/249/base
2025-12-04T08:53:09.5242372Z  * [new branch]                gh/XuehaiPan/249/head   -> origin/gh/XuehaiPan/249/head
2025-12-04T08:53:09.5242563Z  * [new branch]                gh/XuehaiPan/249/orig   -> origin/gh/XuehaiPan/249/orig
2025-12-04T08:53:09.5242747Z  * [new branch]                gh/XuehaiPan/253/base   -> origin/gh/XuehaiPan/253/base
2025-12-04T08:53:09.5242930Z  * [new branch]                gh/XuehaiPan/253/head   -> origin/gh/XuehaiPan/253/head
2025-12-04T08:53:09.5243114Z  * [new branch]                gh/XuehaiPan/253/orig   -> origin/gh/XuehaiPan/253/orig
2025-12-04T08:53:09.5243298Z  * [new branch]                gh/XuehaiPan/254/base   -> origin/gh/XuehaiPan/254/base
2025-12-04T08:53:09.5243484Z  * [new branch]                gh/XuehaiPan/254/head   -> origin/gh/XuehaiPan/254/head
2025-12-04T08:53:09.5243669Z  * [new branch]                gh/XuehaiPan/254/orig   -> origin/gh/XuehaiPan/254/orig
2025-12-04T08:53:09.5243860Z  * [new branch]                gh/XuehaiPan/255/base   -> origin/gh/XuehaiPan/255/base
2025-12-04T08:53:09.5244043Z  * [new branch]                gh/XuehaiPan/255/head   -> origin/gh/XuehaiPan/255/head
2025-12-04T08:53:09.5244228Z  * [new branch]                gh/XuehaiPan/255/orig   -> origin/gh/XuehaiPan/255/orig
2025-12-04T08:53:09.5244412Z  * [new branch]                gh/XuehaiPan/271/base   -> origin/gh/XuehaiPan/271/base
2025-12-04T08:53:09.5244599Z  * [new branch]                gh/XuehaiPan/271/head   -> origin/gh/XuehaiPan/271/head
2025-12-04T08:53:09.5244784Z  * [new branch]                gh/XuehaiPan/271/orig   -> origin/gh/XuehaiPan/271/orig
2025-12-04T08:53:09.5244967Z  * [new branch]                gh/XuehaiPan/343/base   -> origin/gh/XuehaiPan/343/base
2025-12-04T08:53:09.5245209Z  * [new branch]                gh/XuehaiPan/343/head   -> origin/gh/XuehaiPan/343/head
2025-12-04T08:53:09.5245423Z  * [new branch]                gh/XuehaiPan/343/orig   -> origin/gh/XuehaiPan/343/orig
2025-12-04T08:53:09.5245608Z  * [new branch]                gh/XuehaiPan/347/base   -> origin/gh/XuehaiPan/347/base
2025-12-04T08:53:09.5245796Z  * [new branch]                gh/XuehaiPan/347/head   -> origin/gh/XuehaiPan/347/head
2025-12-04T08:53:09.5245982Z  * [new branch]                gh/XuehaiPan/347/orig   -> origin/gh/XuehaiPan/347/orig
2025-12-04T08:53:09.5246166Z  * [new branch]                gh/XuehaiPan/348/base   -> origin/gh/XuehaiPan/348/base
2025-12-04T08:53:09.5246350Z  * [new branch]                gh/XuehaiPan/348/head   -> origin/gh/XuehaiPan/348/head
2025-12-04T08:53:09.5246534Z  * [new branch]                gh/XuehaiPan/348/orig   -> origin/gh/XuehaiPan/348/orig
2025-12-04T08:53:09.5246720Z  * [new branch]                gh/XuehaiPan/350/base   -> origin/gh/XuehaiPan/350/base
2025-12-04T08:53:09.5246916Z  * [new branch]                gh/XuehaiPan/350/head   -> origin/gh/XuehaiPan/350/head
2025-12-04T08:53:09.5247109Z  * [new branch]                gh/XuehaiPan/350/orig   -> origin/gh/XuehaiPan/350/orig
2025-12-04T08:53:09.5247298Z  * [new branch]                gh/XuehaiPan/365/base   -> origin/gh/XuehaiPan/365/base
2025-12-04T08:53:09.5247486Z  * [new branch]                gh/XuehaiPan/365/head   -> origin/gh/XuehaiPan/365/head
2025-12-04T08:53:09.5247674Z  * [new branch]                gh/XuehaiPan/365/orig   -> origin/gh/XuehaiPan/365/orig
2025-12-04T08:53:09.5247861Z  * [new branch]                gh/XuehaiPan/366/base   -> origin/gh/XuehaiPan/366/base
2025-12-04T08:53:09.5248050Z  * [new branch]                gh/XuehaiPan/366/head   -> origin/gh/XuehaiPan/366/head
2025-12-04T08:53:09.5248237Z  * [new branch]                gh/XuehaiPan/370/base   -> origin/gh/XuehaiPan/370/base
2025-12-04T08:53:09.5248423Z  * [new branch]                gh/XuehaiPan/370/head   -> origin/gh/XuehaiPan/370/head
2025-12-04T08:53:09.5248615Z  * [new branch]                gh/XuehaiPan/370/orig   -> origin/gh/XuehaiPan/370/orig
2025-12-04T08:53:09.5248810Z  * [new branch]                gh/XuehaiPan/390/base   -> origin/gh/XuehaiPan/390/base
2025-12-04T08:53:09.5248995Z  * [new branch]                gh/XuehaiPan/390/head   -> origin/gh/XuehaiPan/390/head
2025-12-04T08:53:09.5249185Z  * [new branch]                gh/XuehaiPan/390/orig   -> origin/gh/XuehaiPan/390/orig
2025-12-04T08:53:09.5249374Z  * [new branch]                gh/XuehaiPan/391/base   -> origin/gh/XuehaiPan/391/base
2025-12-04T08:53:09.5249559Z  * [new branch]                gh/XuehaiPan/391/head   -> origin/gh/XuehaiPan/391/head
2025-12-04T08:53:09.5249747Z  * [new branch]                gh/XuehaiPan/391/orig   -> origin/gh/XuehaiPan/391/orig
2025-12-04T08:53:09.5249936Z  * [new branch]                gh/XuehaiPan/392/base   -> origin/gh/XuehaiPan/392/base
2025-12-04T08:53:09.5250120Z  * [new branch]                gh/XuehaiPan/392/head   -> origin/gh/XuehaiPan/392/head
2025-12-04T08:53:09.5250309Z  * [new branch]                gh/XuehaiPan/392/orig   -> origin/gh/XuehaiPan/392/orig
2025-12-04T08:53:09.5250518Z  * [new branch]                gh/XuehaiPan/394/base   -> origin/gh/XuehaiPan/394/base
2025-12-04T08:53:09.5250707Z  * [new branch]                gh/XuehaiPan/394/head   -> origin/gh/XuehaiPan/394/head
2025-12-04T08:53:09.5250894Z  * [new branch]                gh/XuehaiPan/394/orig   -> origin/gh/XuehaiPan/394/orig
2025-12-04T08:53:09.5251078Z  * [new branch]                gh/XuehaiPan/397/base   -> origin/gh/XuehaiPan/397/base
2025-12-04T08:53:09.5251265Z  * [new branch]                gh/XuehaiPan/397/head   -> origin/gh/XuehaiPan/397/head
2025-12-04T08:53:09.5251453Z  * [new branch]                gh/XuehaiPan/397/orig   -> origin/gh/XuehaiPan/397/orig
2025-12-04T08:53:09.5251638Z  * [new branch]                gh/XuehaiPan/398/base   -> origin/gh/XuehaiPan/398/base
2025-12-04T08:53:09.5251865Z  * [new branch]                gh/XuehaiPan/398/head   -> origin/gh/XuehaiPan/398/head
2025-12-04T08:53:09.5252053Z  * [new branch]                gh/XuehaiPan/398/orig   -> origin/gh/XuehaiPan/398/orig
2025-12-04T08:53:09.5252278Z  * [new branch]                gh/XuehaiPan/399/base   -> origin/gh/XuehaiPan/399/base
2025-12-04T08:53:09.5252469Z  * [new branch]                gh/XuehaiPan/399/head   -> origin/gh/XuehaiPan/399/head
2025-12-04T08:53:09.5252659Z  * [new branch]                gh/XuehaiPan/399/orig   -> origin/gh/XuehaiPan/399/orig
2025-12-04T08:53:09.5252845Z  * [new branch]                gh/XuehaiPan/400/base   -> origin/gh/XuehaiPan/400/base
2025-12-04T08:53:09.5253034Z  * [new branch]                gh/XuehaiPan/400/head   -> origin/gh/XuehaiPan/400/head
2025-12-04T08:53:09.5253223Z  * [new branch]                gh/XuehaiPan/400/orig   -> origin/gh/XuehaiPan/400/orig
2025-12-04T08:53:09.5253418Z  * [new branch]                gh/ZhiweiYan-96/39/base -> origin/gh/ZhiweiYan-96/39/base
2025-12-04T08:53:09.5253616Z  * [new branch]                gh/ZhiweiYan-96/39/head -> origin/gh/ZhiweiYan-96/39/head
2025-12-04T08:53:09.5253808Z  * [new branch]                gh/ZhiweiYan-96/39/orig -> origin/gh/ZhiweiYan-96/39/orig
2025-12-04T08:53:09.5254002Z  * [new branch]                gh/ZhiweiYan-96/44/base -> origin/gh/ZhiweiYan-96/44/base
2025-12-04T08:53:09.5254193Z  * [new branch]                gh/ZhiweiYan-96/44/head -> origin/gh/ZhiweiYan-96/44/head
2025-12-04T08:53:09.5254383Z  * [new branch]                gh/ZhiweiYan-96/45/base -> origin/gh/ZhiweiYan-96/45/base
2025-12-04T08:53:09.5254571Z  * [new branch]                gh/ZhiweiYan-96/45/head -> origin/gh/ZhiweiYan-96/45/head
2025-12-04T08:53:09.5254761Z  * [new branch]                gh/ZhiweiYan-96/49/base -> origin/gh/ZhiweiYan-96/49/base
2025-12-04T08:53:09.5254949Z  * [new branch]                gh/ZhiweiYan-96/49/head -> origin/gh/ZhiweiYan-96/49/head
2025-12-04T08:53:09.5255140Z  * [new branch]                gh/ZhiweiYan-96/62/base -> origin/gh/ZhiweiYan-96/62/base
2025-12-04T08:53:09.5255332Z  * [new branch]                gh/ZhiweiYan-96/62/head -> origin/gh/ZhiweiYan-96/62/head
2025-12-04T08:53:09.5255523Z  * [new branch]                gh/ZhiweiYan-96/66/base -> origin/gh/ZhiweiYan-96/66/base
2025-12-04T08:53:09.5255716Z  * [new branch]                gh/ZhiweiYan-96/66/head -> origin/gh/ZhiweiYan-96/66/head
2025-12-04T08:53:09.5255908Z  * [new branch]                gh/ZhiweiYan-96/67/base -> origin/gh/ZhiweiYan-96/67/base
2025-12-04T08:53:09.5256095Z  * [new branch]                gh/ZhiweiYan-96/67/head -> origin/gh/ZhiweiYan-96/67/head
2025-12-04T08:53:09.5256285Z  * [new branch]                gh/ZhiweiYan-96/68/base -> origin/gh/ZhiweiYan-96/68/base
2025-12-04T08:53:09.5256477Z  * [new branch]                gh/ZhiweiYan-96/68/head -> origin/gh/ZhiweiYan-96/68/head
2025-12-04T08:53:09.5256664Z  * [new branch]                gh/ZhiweiYan-96/68/orig -> origin/gh/ZhiweiYan-96/68/orig
2025-12-04T08:53:09.5256852Z  * [new branch]                gh/aakhundov/1/base     -> origin/gh/aakhundov/1/base
2025-12-04T08:53:09.5257041Z  * [new branch]                gh/aakhundov/1/head     -> origin/gh/aakhundov/1/head
2025-12-04T08:53:09.5257227Z  * [new branch]                gh/aakhundov/2/base     -> origin/gh/aakhundov/2/base
2025-12-04T08:53:09.5257411Z  * [new branch]                gh/aakhundov/2/head     -> origin/gh/aakhundov/2/head
2025-12-04T08:53:09.5257598Z  * [new branch]                gh/aditew01/openblas    -> origin/gh/aditew01/openblas
2025-12-04T08:53:09.5257785Z  * [new branch]                gh/aditew01/sbgemm      -> origin/gh/aditew01/sbgemm
2025-12-04T08:53:09.5257970Z  * [new branch]                gh/aditew01/vecbf16     -> origin/gh/aditew01/vecbf16
2025-12-04T08:53:09.5258150Z  * [new branch]                gh/albanD/4/base        -> origin/gh/albanD/4/base
2025-12-04T08:53:09.5258323Z  * [new branch]                gh/albanD/4/head        -> origin/gh/albanD/4/head
2025-12-04T08:53:09.5258498Z  * [new branch]                gh/albanD/4/orig        -> origin/gh/albanD/4/orig
2025-12-04T08:53:09.5258798Z  * [new branch]                gh/alexbrauckmann/paddedtensor_faketensor_init -> origin/gh/alexbrauckmann/paddedtensor_faketensor_init
2025-12-04T08:53:09.5259100Z  * [new branch]                gh/alexsamardzic/12/base -> origin/gh/alexsamardzic/12/base
2025-12-04T08:53:09.5259305Z  * [new branch]                gh/alexsamardzic/12/head -> origin/gh/alexsamardzic/12/head
2025-12-04T08:53:09.5259508Z  * [new branch]                gh/alexsamardzic/12/orig -> origin/gh/alexsamardzic/12/orig
2025-12-04T08:53:09.5259705Z  * [new branch]                gh/alexsamardzic/14/base -> origin/gh/alexsamardzic/14/base
2025-12-04T08:53:09.5259904Z  * [new branch]                gh/alexsamardzic/14/head -> origin/gh/alexsamardzic/14/head
2025-12-04T08:53:09.5260101Z  * [new branch]                gh/alexsamardzic/14/orig -> origin/gh/alexsamardzic/14/orig
2025-12-04T08:53:09.5260300Z  * [new branch]                gh/alexsamardzic/15/base -> origin/gh/alexsamardzic/15/base
2025-12-04T08:53:09.5260544Z  * [new branch]                gh/alexsamardzic/15/head -> origin/gh/alexsamardzic/15/head
2025-12-04T08:53:09.5260747Z  * [new branch]                gh/alexsamardzic/15/orig -> origin/gh/alexsamardzic/15/orig
2025-12-04T08:53:09.5260940Z  * [new branch]                gh/amjames/18/base      -> origin/gh/amjames/18/base
2025-12-04T08:53:09.5261124Z  * [new branch]                gh/amjames/18/head      -> origin/gh/amjames/18/head
2025-12-04T08:53:09.5261303Z  * [new branch]                gh/amjames/18/orig      -> origin/gh/amjames/18/orig
2025-12-04T08:53:09.5261489Z  * [new branch]                gh/andrewor14/35/base   -> origin/gh/andrewor14/35/base
2025-12-04T08:53:09.5261681Z  * [new branch]                gh/andrewor14/35/head   -> origin/gh/andrewor14/35/head
2025-12-04T08:53:09.5261866Z  * [new branch]                gh/andrewor14/35/orig   -> origin/gh/andrewor14/35/orig
2025-12-04T08:53:09.5262053Z  * [new branch]                gh/andrewor14/50/base   -> origin/gh/andrewor14/50/base
2025-12-04T08:53:09.5262244Z  * [new branch]                gh/andrewor14/50/head   -> origin/gh/andrewor14/50/head
2025-12-04T08:53:09.5262429Z  * [new branch]                gh/andrewor14/50/orig   -> origin/gh/andrewor14/50/orig
2025-12-04T08:53:09.5262615Z  * [new branch]                gh/andyanwang/30/base   -> origin/gh/andyanwang/30/base
2025-12-04T08:53:09.5262804Z  * [new branch]                gh/andyanwang/30/orig   -> origin/gh/andyanwang/30/orig
2025-12-04T08:53:09.5262989Z  * [new branch]                gh/andyanwang/31/base   -> origin/gh/andyanwang/31/base
2025-12-04T08:53:09.5263178Z  * [new branch]                gh/andyanwang/31/orig   -> origin/gh/andyanwang/31/orig
2025-12-04T08:53:09.5263367Z  * [new branch]                gh/andyanwang/39/base   -> origin/gh/andyanwang/39/base
2025-12-04T08:53:09.5263552Z  * [new branch]                gh/andyanwang/39/head   -> origin/gh/andyanwang/39/head
2025-12-04T08:53:09.5263742Z  * [new branch]                gh/andyanwang/39/orig   -> origin/gh/andyanwang/39/orig
2025-12-04T08:53:09.5263930Z  * [new branch]                gh/andyanwang/42/base   -> origin/gh/andyanwang/42/base
2025-12-04T08:53:09.5264120Z  * [new branch]                gh/andyanwang/42/head   -> origin/gh/andyanwang/42/head
2025-12-04T08:53:09.5264307Z  * [new branch]                gh/andyanwang/42/orig   -> origin/gh/andyanwang/42/orig
2025-12-04T08:53:09.5264496Z  * [new branch]                gh/andyanwang/45/base   -> origin/gh/andyanwang/45/base
2025-12-04T08:53:09.5264685Z  * [new branch]                gh/andyanwang/45/head   -> origin/gh/andyanwang/45/head
2025-12-04T08:53:09.5264873Z  * [new branch]                gh/andyanwang/45/orig   -> origin/gh/andyanwang/45/orig
2025-12-04T08:53:09.5265059Z  * [new branch]                gh/angelayi/107/base    -> origin/gh/angelayi/107/base
2025-12-04T08:53:09.5265249Z  * [new branch]                gh/angelayi/107/head    -> origin/gh/angelayi/107/head
2025-12-04T08:53:09.5265477Z  * [new branch]                gh/angelayi/114/base    -> origin/gh/angelayi/114/base
2025-12-04T08:53:09.5265658Z  * [new branch]                gh/angelayi/114/head    -> origin/gh/angelayi/114/head
2025-12-04T08:53:09.5265874Z  * [new branch]                gh/angelayi/114/orig    -> origin/gh/angelayi/114/orig
2025-12-04T08:53:09.5266058Z  * [new branch]                gh/angelayi/116/base    -> origin/gh/angelayi/116/base
2025-12-04T08:53:09.5266239Z  * [new branch]                gh/angelayi/116/head    -> origin/gh/angelayi/116/head
2025-12-04T08:53:09.5266424Z  * [new branch]                gh/angelayi/116/orig    -> origin/gh/angelayi/116/orig
2025-12-04T08:53:09.5266608Z  * [new branch]                gh/angelayi/122/base    -> origin/gh/angelayi/122/base
2025-12-04T08:53:09.5266789Z  * [new branch]                gh/angelayi/122/head    -> origin/gh/angelayi/122/head
2025-12-04T08:53:09.5266973Z  * [new branch]                gh/angelayi/122/orig    -> origin/gh/angelayi/122/orig
2025-12-04T08:53:09.5267163Z  * [new branch]                gh/angelayi/124/base    -> origin/gh/angelayi/124/base
2025-12-04T08:53:09.5267346Z  * [new branch]                gh/angelayi/124/head    -> origin/gh/angelayi/124/head
2025-12-04T08:53:09.5267537Z  * [new branch]                gh/angelayi/124/orig    -> origin/gh/angelayi/124/orig
2025-12-04T08:53:09.5267722Z  * [new branch]                gh/angelayi/128/base    -> origin/gh/angelayi/128/base
2025-12-04T08:53:09.5267906Z  * [new branch]                gh/angelayi/128/head    -> origin/gh/angelayi/128/head
2025-12-04T08:53:09.5268091Z  * [new branch]                gh/angelayi/128/orig    -> origin/gh/angelayi/128/orig
2025-12-04T08:53:09.5268274Z  * [new branch]                gh/angelayi/131/base    -> origin/gh/angelayi/131/base
2025-12-04T08:53:09.5268456Z  * [new branch]                gh/angelayi/131/head    -> origin/gh/angelayi/131/head
2025-12-04T08:53:09.5268640Z  * [new branch]                gh/angelayi/131/orig    -> origin/gh/angelayi/131/orig
2025-12-04T08:53:09.5268827Z  * [new branch]                gh/angelayi/132/base    -> origin/gh/angelayi/132/base
2025-12-04T08:53:09.5269012Z  * [new branch]                gh/angelayi/132/head    -> origin/gh/angelayi/132/head
2025-12-04T08:53:09.5269196Z  * [new branch]                gh/angelayi/132/orig    -> origin/gh/angelayi/132/orig
2025-12-04T08:53:09.5269378Z  * [new branch]                gh/angelayi/133/base    -> origin/gh/angelayi/133/base
2025-12-04T08:53:09.5269562Z  * [new branch]                gh/angelayi/133/head    -> origin/gh/angelayi/133/head
2025-12-04T08:53:09.5269747Z  * [new branch]                gh/angelayi/133/orig    -> origin/gh/angelayi/133/orig
2025-12-04T08:53:09.5269929Z  * [new branch]                gh/angelayi/134/base    -> origin/gh/angelayi/134/base
2025-12-04T08:53:09.5270114Z  * [new branch]                gh/angelayi/134/head    -> origin/gh/angelayi/134/head
2025-12-04T08:53:09.5270299Z  * [new branch]                gh/angelayi/134/orig    -> origin/gh/angelayi/134/orig
2025-12-04T08:53:09.5270515Z  * [new branch]                gh/angelayi/135/base    -> origin/gh/angelayi/135/base
2025-12-04T08:53:09.5270702Z  * [new branch]                gh/angelayi/135/head    -> origin/gh/angelayi/135/head
2025-12-04T08:53:09.5270887Z  * [new branch]                gh/angelayi/135/orig    -> origin/gh/angelayi/135/orig
2025-12-04T08:53:09.5271071Z  * [new branch]                gh/angelayi/136/base    -> origin/gh/angelayi/136/base
2025-12-04T08:53:09.5271256Z  * [new branch]                gh/angelayi/136/head    -> origin/gh/angelayi/136/head
2025-12-04T08:53:09.5271442Z  * [new branch]                gh/angelayi/136/orig    -> origin/gh/angelayi/136/orig
2025-12-04T08:53:09.5271623Z  * [new branch]                gh/angelayi/137/base    -> origin/gh/angelayi/137/base
2025-12-04T08:53:09.5271806Z  * [new branch]                gh/angelayi/137/head    -> origin/gh/angelayi/137/head
2025-12-04T08:53:09.5271989Z  * [new branch]                gh/angelayi/137/orig    -> origin/gh/angelayi/137/orig
2025-12-04T08:53:09.5272210Z  * [new branch]                gh/angelayi/138/base    -> origin/gh/angelayi/138/base
2025-12-04T08:53:09.5272425Z  * [new branch]                gh/angelayi/138/head    -> origin/gh/angelayi/138/head
2025-12-04T08:53:09.5272607Z  * [new branch]                gh/angelayi/138/orig    -> origin/gh/angelayi/138/orig
2025-12-04T08:53:09.5272791Z  * [new branch]                gh/angelayi/139/base    -> origin/gh/angelayi/139/base
2025-12-04T08:53:09.5272973Z  * [new branch]                gh/angelayi/139/head    -> origin/gh/angelayi/139/head
2025-12-04T08:53:09.5273153Z  * [new branch]                gh/angelayi/139/orig    -> origin/gh/angelayi/139/orig
2025-12-04T08:53:09.5273334Z  * [new branch]                gh/angelayi/140/base    -> origin/gh/angelayi/140/base
2025-12-04T08:53:09.5273521Z  * [new branch]                gh/angelayi/140/head    -> origin/gh/angelayi/140/head
2025-12-04T08:53:09.5273703Z  * [new branch]                gh/angelayi/140/orig    -> origin/gh/angelayi/140/orig
2025-12-04T08:53:09.5273891Z  * [new branch]                gh/angelayi/141/base    -> origin/gh/angelayi/141/base
2025-12-04T08:53:09.5274080Z  * [new branch]                gh/angelayi/141/head    -> origin/gh/angelayi/141/head
2025-12-04T08:53:09.5274261Z  * [new branch]                gh/angelayi/141/orig    -> origin/gh/angelayi/141/orig
2025-12-04T08:53:09.5274447Z  * [new branch]                gh/angelayi/142/base    -> origin/gh/angelayi/142/base
2025-12-04T08:53:09.5274632Z  * [new branch]                gh/angelayi/142/head    -> origin/gh/angelayi/142/head
2025-12-04T08:53:09.5274813Z  * [new branch]                gh/angelayi/142/orig    -> origin/gh/angelayi/142/orig
2025-12-04T08:53:09.5274997Z  * [new branch]                gh/angelayi/143/base    -> origin/gh/angelayi/143/base
2025-12-04T08:53:09.5275185Z  * [new branch]                gh/angelayi/143/head    -> origin/gh/angelayi/143/head
2025-12-04T08:53:09.5275447Z  * [new branch]                gh/angelayi/143/orig    -> origin/gh/angelayi/143/orig
2025-12-04T08:53:09.5275637Z  * [new branch]                gh/angelayi/144/base    -> origin/gh/angelayi/144/base
2025-12-04T08:53:09.5275826Z  * [new branch]                gh/angelayi/144/head    -> origin/gh/angelayi/144/head
2025-12-04T08:53:09.5276005Z  * [new branch]                gh/angelayi/144/orig    -> origin/gh/angelayi/144/orig
2025-12-04T08:53:09.5276194Z  * [new branch]                gh/anijain2305/753/base -> origin/gh/anijain2305/753/base
2025-12-04T08:53:09.5276389Z  * [new branch]                gh/anijain2305/753/head -> origin/gh/anijain2305/753/head
2025-12-04T08:53:09.5276579Z  * [new branch]                gh/anijain2305/753/orig -> origin/gh/anijain2305/753/orig
2025-12-04T08:53:09.5276768Z  * [new branch]                gh/anijain2305/810/base -> origin/gh/anijain2305/810/base
2025-12-04T08:53:09.5276957Z  * [new branch]                gh/anijain2305/810/head -> origin/gh/anijain2305/810/head
2025-12-04T08:53:09.5277149Z  * [new branch]                gh/anijain2305/810/orig -> origin/gh/anijain2305/810/orig
2025-12-04T08:53:09.5277337Z  * [new branch]                gh/anijain2305/854/base -> origin/gh/anijain2305/854/base
2025-12-04T08:53:09.5277525Z  * [new branch]                gh/anijain2305/854/head -> origin/gh/anijain2305/854/head
2025-12-04T08:53:09.5277714Z  * [new branch]                gh/anijain2305/854/orig -> origin/gh/anijain2305/854/orig
2025-12-04T08:53:09.5277901Z  * [new branch]                gh/anijain2305/864/base -> origin/gh/anijain2305/864/base
2025-12-04T08:53:09.5278088Z  * [new branch]                gh/anijain2305/864/head -> origin/gh/anijain2305/864/head
2025-12-04T08:53:09.5278274Z  * [new branch]                gh/anijain2305/864/orig -> origin/gh/anijain2305/864/orig
2025-12-04T08:53:09.5278463Z  * [new branch]                gh/anijain2305/870/base -> origin/gh/anijain2305/870/base
2025-12-04T08:53:09.5278653Z  * [new branch]                gh/anijain2305/870/head -> origin/gh/anijain2305/870/head
2025-12-04T08:53:09.5278873Z  * [new branch]                gh/anijain2305/870/orig -> origin/gh/anijain2305/870/orig
2025-12-04T08:53:09.5279094Z  * [new branch]                gh/anijain2305/873/base -> origin/gh/anijain2305/873/base
2025-12-04T08:53:09.5279281Z  * [new branch]                gh/anijain2305/873/head -> origin/gh/anijain2305/873/head
2025-12-04T08:53:09.5279470Z  * [new branch]                gh/anijain2305/873/orig -> origin/gh/anijain2305/873/orig
2025-12-04T08:53:09.5279657Z  * [new branch]                gh/anijain2305/894/base -> origin/gh/anijain2305/894/base
2025-12-04T08:53:09.5279843Z  * [new branch]                gh/anijain2305/894/head -> origin/gh/anijain2305/894/head
2025-12-04T08:53:09.5280030Z  * [new branch]                gh/anijain2305/894/orig -> origin/gh/anijain2305/894/orig
2025-12-04T08:53:09.5280216Z  * [new branch]                gh/anijain2305/895/base -> origin/gh/anijain2305/895/base
2025-12-04T08:53:09.5280401Z  * [new branch]                gh/anijain2305/895/head -> origin/gh/anijain2305/895/head
2025-12-04T08:53:09.5280631Z  * [new branch]                gh/anijain2305/895/orig -> origin/gh/anijain2305/895/orig
2025-12-04T08:53:09.5280824Z  * [new branch]                gh/anijain2305/910/base -> origin/gh/anijain2305/910/base
2025-12-04T08:53:09.5281010Z  * [new branch]                gh/anijain2305/910/head -> origin/gh/anijain2305/910/head
2025-12-04T08:53:09.5281202Z  * [new branch]                gh/anijain2305/910/orig -> origin/gh/anijain2305/910/orig
2025-12-04T08:53:09.5281392Z  * [new branch]                gh/anijain2305/919/base -> origin/gh/anijain2305/919/base
2025-12-04T08:53:09.5281578Z  * [new branch]                gh/anijain2305/919/head -> origin/gh/anijain2305/919/head
2025-12-04T08:53:09.5281767Z  * [new branch]                gh/anijain2305/919/orig -> origin/gh/anijain2305/919/orig
2025-12-04T08:53:09.5281956Z  * [new branch]                gh/anijain2305/922/base -> origin/gh/anijain2305/922/base
2025-12-04T08:53:09.5282146Z  * [new branch]                gh/anijain2305/922/head -> origin/gh/anijain2305/922/head
2025-12-04T08:53:09.5282339Z  * [new branch]                gh/anijain2305/922/orig -> origin/gh/anijain2305/922/orig
2025-12-04T08:53:09.5282533Z  * [new branch]                gh/anijain2305/932/base -> origin/gh/anijain2305/932/base
2025-12-04T08:53:09.5282719Z  * [new branch]                gh/anijain2305/932/head -> origin/gh/anijain2305/932/head
2025-12-04T08:53:09.5282915Z  * [new branch]                gh/anijain2305/932/orig -> origin/gh/anijain2305/932/orig
2025-12-04T08:53:09.5283103Z  * [new branch]                gh/anijain2305/940/base -> origin/gh/anijain2305/940/base
2025-12-04T08:53:09.5283290Z  * [new branch]                gh/anijain2305/940/head -> origin/gh/anijain2305/940/head
2025-12-04T08:53:09.5283480Z  * [new branch]                gh/anijain2305/940/orig -> origin/gh/anijain2305/940/orig
2025-12-04T08:53:09.5283666Z  * [new branch]                gh/anijain2305/941/base -> origin/gh/anijain2305/941/base
2025-12-04T08:53:09.5283863Z  * [new branch]                gh/anijain2305/941/head -> origin/gh/anijain2305/941/head
2025-12-04T08:53:09.5284053Z  * [new branch]                gh/anijain2305/941/orig -> origin/gh/anijain2305/941/orig
2025-12-04T08:53:09.5284244Z  * [new branch]                gh/anijain2305/942/base -> origin/gh/anijain2305/942/base
2025-12-04T08:53:09.5284434Z  * [new branch]                gh/anijain2305/942/head -> origin/gh/anijain2305/942/head
2025-12-04T08:53:09.5284625Z  * [new branch]                gh/anijain2305/942/orig -> origin/gh/anijain2305/942/orig
2025-12-04T08:53:09.5284813Z  * [new branch]                gh/anijain2305/943/base -> origin/gh/anijain2305/943/base
2025-12-04T08:53:09.5285000Z  * [new branch]                gh/anijain2305/943/head -> origin/gh/anijain2305/943/head
2025-12-04T08:53:09.5285187Z  * [new branch]                gh/anijain2305/943/orig -> origin/gh/anijain2305/943/orig
2025-12-04T08:53:09.5285374Z  * [new branch]                gh/anijain2305/944/base -> origin/gh/anijain2305/944/base
2025-12-04T08:53:09.5285622Z  * [new branch]                gh/anijain2305/944/head -> origin/gh/anijain2305/944/head
2025-12-04T08:53:09.5285858Z  * [new branch]                gh/anijain2305/944/orig -> origin/gh/anijain2305/944/orig
2025-12-04T08:53:09.5286045Z  * [new branch]                gh/anijain2305/945/base -> origin/gh/anijain2305/945/base
2025-12-04T08:53:09.5286233Z  * [new branch]                gh/anijain2305/945/head -> origin/gh/anijain2305/945/head
2025-12-04T08:53:09.5286420Z  * [new branch]                gh/anijain2305/945/orig -> origin/gh/anijain2305/945/orig
2025-12-04T08:53:09.5286606Z  * [new branch]                gh/anijain2305/946/base -> origin/gh/anijain2305/946/base
2025-12-04T08:53:09.5286800Z  * [new branch]                gh/anijain2305/946/head -> origin/gh/anijain2305/946/head
2025-12-04T08:53:09.5286990Z  * [new branch]                gh/anijain2305/946/orig -> origin/gh/anijain2305/946/orig
2025-12-04T08:53:09.5287178Z  * [new branch]                gh/anijain2305/947/base -> origin/gh/anijain2305/947/base
2025-12-04T08:53:09.5287373Z  * [new branch]                gh/anijain2305/947/head -> origin/gh/anijain2305/947/head
2025-12-04T08:53:09.5287561Z  * [new branch]                gh/anijain2305/947/orig -> origin/gh/anijain2305/947/orig
2025-12-04T08:53:09.5287750Z  * [new branch]                gh/anijain2305/948/base -> origin/gh/anijain2305/948/base
2025-12-04T08:53:09.5287941Z  * [new branch]                gh/anijain2305/948/head -> origin/gh/anijain2305/948/head
2025-12-04T08:53:09.5288128Z  * [new branch]                gh/anijain2305/948/orig -> origin/gh/anijain2305/948/orig
2025-12-04T08:53:09.5288316Z  * [new branch]                gh/anijain2305/949/base -> origin/gh/anijain2305/949/base
2025-12-04T08:53:09.5288506Z  * [new branch]                gh/anijain2305/949/head -> origin/gh/anijain2305/949/head
2025-12-04T08:53:09.5288692Z  * [new branch]                gh/anijain2305/949/orig -> origin/gh/anijain2305/949/orig
2025-12-04T08:53:09.5288886Z  * [new branch]                gh/anijain2305/950/base -> origin/gh/anijain2305/950/base
2025-12-04T08:53:09.5289080Z  * [new branch]                gh/anijain2305/950/head -> origin/gh/anijain2305/950/head
2025-12-04T08:53:09.5289269Z  * [new branch]                gh/anijain2305/950/orig -> origin/gh/anijain2305/950/orig
2025-12-04T08:53:09.5289459Z  * [new branch]                gh/anijain2305/951/base -> origin/gh/anijain2305/951/base
2025-12-04T08:53:09.5289648Z  * [new branch]                gh/anijain2305/951/head -> origin/gh/anijain2305/951/head
2025-12-04T08:53:09.5289838Z  * [new branch]                gh/anijain2305/951/orig -> origin/gh/anijain2305/951/orig
2025-12-04T08:53:09.5290026Z  * [new branch]                gh/anijain2305/952/base -> origin/gh/anijain2305/952/base
2025-12-04T08:53:09.5290214Z  * [new branch]                gh/anijain2305/952/head -> origin/gh/anijain2305/952/head
2025-12-04T08:53:09.5290400Z  * [new branch]                gh/anijain2305/952/orig -> origin/gh/anijain2305/952/orig
2025-12-04T08:53:09.5290620Z  * [new branch]                gh/anijain2305/953/base -> origin/gh/anijain2305/953/base
2025-12-04T08:53:09.5290814Z  * [new branch]                gh/anijain2305/953/head -> origin/gh/anijain2305/953/head
2025-12-04T08:53:09.5290999Z  * [new branch]                gh/anijain2305/953/orig -> origin/gh/anijain2305/953/orig
2025-12-04T08:53:09.5291188Z  * [new branch]                gh/anijain2305/954/base -> origin/gh/anijain2305/954/base
2025-12-04T08:53:09.5291380Z  * [new branch]                gh/anijain2305/954/head -> origin/gh/anijain2305/954/head
2025-12-04T08:53:09.5291567Z  * [new branch]                gh/anijain2305/954/orig -> origin/gh/anijain2305/954/orig
2025-12-04T08:53:09.5291755Z  * [new branch]                gh/anijain2305/955/base -> origin/gh/anijain2305/955/base
2025-12-04T08:53:09.5291942Z  * [new branch]                gh/anijain2305/955/head -> origin/gh/anijain2305/955/head
2025-12-04T08:53:09.5292128Z  * [new branch]                gh/anijain2305/955/orig -> origin/gh/anijain2305/955/orig
2025-12-04T08:53:09.5292370Z  * [new branch]                gh/anijain2305/956/base -> origin/gh/anijain2305/956/base
2025-12-04T08:53:09.5292757Z  * [new branch]                gh/anijain2305/956/head -> origin/gh/anijain2305/956/head
2025-12-04T08:53:09.5292949Z  * [new branch]                gh/anijain2305/956/orig -> origin/gh/anijain2305/956/orig
2025-12-04T08:53:09.5293136Z  * [new branch]                gh/anijain2305/957/base -> origin/gh/anijain2305/957/base
2025-12-04T08:53:09.5293323Z  * [new branch]                gh/anijain2305/957/head -> origin/gh/anijain2305/957/head
2025-12-04T08:53:09.5293515Z  * [new branch]                gh/anijain2305/957/orig -> origin/gh/anijain2305/957/orig
2025-12-04T08:53:09.5293702Z  * [new branch]                gh/anijain2305/958/base -> origin/gh/anijain2305/958/base
2025-12-04T08:53:09.5293888Z  * [new branch]                gh/anijain2305/958/head -> origin/gh/anijain2305/958/head
2025-12-04T08:53:09.5294082Z  * [new branch]                gh/anijain2305/958/orig -> origin/gh/anijain2305/958/orig
2025-12-04T08:53:09.5294274Z  * [new branch]                gh/anijain2305/959/base -> origin/gh/anijain2305/959/base
2025-12-04T08:53:09.5294466Z  * [new branch]                gh/anijain2305/959/head -> origin/gh/anijain2305/959/head
2025-12-04T08:53:09.5294655Z  * [new branch]                gh/anijain2305/959/orig -> origin/gh/anijain2305/959/orig
2025-12-04T08:53:09.5294843Z  * [new branch]                gh/anijain2305/960/base -> origin/gh/anijain2305/960/base
2025-12-04T08:53:09.5295030Z  * [new branch]                gh/anijain2305/960/head -> origin/gh/anijain2305/960/head
2025-12-04T08:53:09.5295223Z  * [new branch]                gh/anijain2305/960/orig -> origin/gh/anijain2305/960/orig
2025-12-04T08:53:09.5295415Z  * [new branch]                gh/anijain2305/961/base -> origin/gh/anijain2305/961/base
2025-12-04T08:53:09.5295603Z  * [new branch]                gh/anijain2305/961/head -> origin/gh/anijain2305/961/head
2025-12-04T08:53:09.5295801Z  * [new branch]                gh/anijain2305/961/orig -> origin/gh/anijain2305/961/orig
2025-12-04T08:53:09.5295996Z  * [new branch]                gh/anijain2305/962/base -> origin/gh/anijain2305/962/base
2025-12-04T08:53:09.5296187Z  * [new branch]                gh/anijain2305/962/head -> origin/gh/anijain2305/962/head
2025-12-04T08:53:09.5296378Z  * [new branch]                gh/anijain2305/962/orig -> origin/gh/anijain2305/962/orig
2025-12-04T08:53:09.5296570Z  * [new branch]                gh/anijain2305/963/base -> origin/gh/anijain2305/963/base
2025-12-04T08:53:09.5296758Z  * [new branch]                gh/anijain2305/963/head -> origin/gh/anijain2305/963/head
2025-12-04T08:53:09.5296952Z  * [new branch]                gh/anijain2305/963/orig -> origin/gh/anijain2305/963/orig
2025-12-04T08:53:09.5297142Z  * [new branch]                gh/anijain2305/964/base -> origin/gh/anijain2305/964/base
2025-12-04T08:53:09.5297337Z  * [new branch]                gh/anijain2305/964/head -> origin/gh/anijain2305/964/head
2025-12-04T08:53:09.5297533Z  * [new branch]                gh/anijain2305/964/orig -> origin/gh/anijain2305/964/orig
2025-12-04T08:53:09.5297723Z  * [new branch]                gh/anijain2305/965/base -> origin/gh/anijain2305/965/base
2025-12-04T08:53:09.5297914Z  * [new branch]                gh/anijain2305/965/head -> origin/gh/anijain2305/965/head
2025-12-04T08:53:09.5298109Z  * [new branch]                gh/anijain2305/965/orig -> origin/gh/anijain2305/965/orig
2025-12-04T08:53:09.5298301Z  * [new branch]                gh/anijain2305/966/base -> origin/gh/anijain2305/966/base
2025-12-04T08:53:09.5298492Z  * [new branch]                gh/anijain2305/966/head -> origin/gh/anijain2305/966/head
2025-12-04T08:53:09.5298688Z  * [new branch]                gh/anijain2305/966/orig -> origin/gh/anijain2305/966/orig
2025-12-04T08:53:09.5298877Z  * [new branch]                gh/anijain2305/967/base -> origin/gh/anijain2305/967/base
2025-12-04T08:53:09.5299108Z  * [new branch]                gh/anijain2305/967/head -> origin/gh/anijain2305/967/head
2025-12-04T08:53:09.5299299Z  * [new branch]                gh/anijain2305/967/orig -> origin/gh/anijain2305/967/orig
2025-12-04T08:53:09.5299518Z  * [new branch]                gh/anijain2305/968/base -> origin/gh/anijain2305/968/base
2025-12-04T08:53:09.5299712Z  * [new branch]                gh/anijain2305/968/head -> origin/gh/anijain2305/968/head
2025-12-04T08:53:09.5299902Z  * [new branch]                gh/anijain2305/968/orig -> origin/gh/anijain2305/968/orig
2025-12-04T08:53:09.5300092Z  * [new branch]                gh/anijain2305/969/base -> origin/gh/anijain2305/969/base
2025-12-04T08:53:09.5300284Z  * [new branch]                gh/anijain2305/969/head -> origin/gh/anijain2305/969/head
2025-12-04T08:53:09.5300517Z  * [new branch]                gh/anijain2305/969/orig -> origin/gh/anijain2305/969/orig
2025-12-04T08:53:09.5300707Z  * [new branch]                gh/anijain2305/970/base -> origin/gh/anijain2305/970/base
2025-12-04T08:53:09.5300903Z  * [new branch]                gh/anijain2305/970/head -> origin/gh/anijain2305/970/head
2025-12-04T08:53:09.5301099Z  * [new branch]                gh/anijain2305/970/orig -> origin/gh/anijain2305/970/orig
2025-12-04T08:53:09.5301289Z  * [new branch]                gh/anjali411/216/base   -> origin/gh/anjali411/216/base
2025-12-04T08:53:09.5301479Z  * [new branch]                gh/anjali411/216/head   -> origin/gh/anjali411/216/head
2025-12-04T08:53:09.5301665Z  * [new branch]                gh/anjali411/216/orig   -> origin/gh/anjali411/216/orig
2025-12-04T08:53:09.5301854Z  * [new branch]                gh/anshul-si/1/base     -> origin/gh/anshul-si/1/base
2025-12-04T08:53:09.5302039Z  * [new branch]                gh/anshul-si/1/head     -> origin/gh/anshul-si/1/head
2025-12-04T08:53:09.5302219Z  * [new branch]                gh/anshul-si/2/base     -> origin/gh/anshul-si/2/base
2025-12-04T08:53:09.5302401Z  * [new branch]                gh/anshul-si/2/head     -> origin/gh/anshul-si/2/head
2025-12-04T08:53:09.5302587Z  * [new branch]                gh/anshul-si/3/base     -> origin/gh/anshul-si/3/base
2025-12-04T08:53:09.5302769Z  * [new branch]                gh/anshul-si/3/head     -> origin/gh/anshul-si/3/head
2025-12-04T08:53:09.5302951Z  * [new branch]                gh/anshul-si/4/base     -> origin/gh/anshul-si/4/base
2025-12-04T08:53:09.5303133Z  * [new branch]                gh/anshul-si/4/head     -> origin/gh/anshul-si/4/head
2025-12-04T08:53:09.5303311Z  * [new branch]                gh/anshul-si/5/base     -> origin/gh/anshul-si/5/base
2025-12-04T08:53:09.5303492Z  * [new branch]                gh/anshul-si/5/head     -> origin/gh/anshul-si/5/head
2025-12-04T08:53:09.5303678Z  * [new branch]                gh/anshul-si/53/base    -> origin/gh/anshul-si/53/base
2025-12-04T08:53:09.5303859Z  * [new branch]                gh/anshul-si/53/head    -> origin/gh/anshul-si/53/head
2025-12-04T08:53:09.5304044Z  * [new branch]                gh/anshul-si/58/base    -> origin/gh/anshul-si/58/base
2025-12-04T08:53:09.5304230Z  * [new branch]                gh/anshul-si/58/head    -> origin/gh/anshul-si/58/head
2025-12-04T08:53:09.5304411Z  * [new branch]                gh/anshul-si/66/base    -> origin/gh/anshul-si/66/base
2025-12-04T08:53:09.5304592Z  * [new branch]                gh/anshul-si/66/head    -> origin/gh/anshul-si/66/head
2025-12-04T08:53:09.5304774Z  * [new branch]                gh/anshul-si/66/orig    -> origin/gh/anshul-si/66/orig
2025-12-04T08:53:09.5304961Z  * [new branch]                gh/anshul-si/67/base    -> origin/gh/anshul-si/67/base
2025-12-04T08:53:09.5305139Z  * [new branch]                gh/anshul-si/67/head    -> origin/gh/anshul-si/67/head
2025-12-04T08:53:09.5305320Z  * [new branch]                gh/anshul-si/67/orig    -> origin/gh/anshul-si/67/orig
2025-12-04T08:53:09.5305647Z  * [new branch]                gh/anshul-si/68/base    -> origin/gh/anshul-si/68/base
2025-12-04T08:53:09.5305824Z  * [new branch]                gh/anshul-si/68/head    -> origin/gh/anshul-si/68/head
2025-12-04T08:53:09.5306059Z  * [new branch]                gh/anshul-si/68/orig    -> origin/gh/anshul-si/68/orig
2025-12-04T08:53:09.5306286Z  * [new branch]                gh/anshul-si/69/base    -> origin/gh/anshul-si/69/base
2025-12-04T08:53:09.5306465Z  * [new branch]                gh/anshul-si/69/head    -> origin/gh/anshul-si/69/head
2025-12-04T08:53:09.5306646Z  * [new branch]                gh/anshul-si/69/orig    -> origin/gh/anshul-si/69/orig
2025-12-04T08:53:09.5306827Z  * [new branch]                gh/anshul-si/70/base    -> origin/gh/anshul-si/70/base
2025-12-04T08:53:09.5307007Z  * [new branch]                gh/anshul-si/70/head    -> origin/gh/anshul-si/70/head
2025-12-04T08:53:09.5307185Z  * [new branch]                gh/anshul-si/70/orig    -> origin/gh/anshul-si/70/orig
2025-12-04T08:53:09.5307364Z  * [new branch]                gh/anshul-si/71/base    -> origin/gh/anshul-si/71/base
2025-12-04T08:53:09.5307545Z  * [new branch]                gh/anshul-si/71/head    -> origin/gh/anshul-si/71/head
2025-12-04T08:53:09.5307732Z  * [new branch]                gh/anshul-si/71/orig    -> origin/gh/anshul-si/71/orig
2025-12-04T08:53:09.5307920Z  * [new branch]                gh/anshul-si/72/base    -> origin/gh/anshul-si/72/base
2025-12-04T08:53:09.5308103Z  * [new branch]                gh/anshul-si/72/head    -> origin/gh/anshul-si/72/head
2025-12-04T08:53:09.5308286Z  * [new branch]                gh/anshul-si/72/orig    -> origin/gh/anshul-si/72/orig
2025-12-04T08:53:09.5308465Z  * [new branch]                gh/anshul-si/73/base    -> origin/gh/anshul-si/73/base
2025-12-04T08:53:09.5308646Z  * [new branch]                gh/anshul-si/73/head    -> origin/gh/anshul-si/73/head
2025-12-04T08:53:09.5308830Z  * [new branch]                gh/anshul-si/73/orig    -> origin/gh/anshul-si/73/orig
2025-12-04T08:53:09.5309009Z  * [new branch]                gh/aorenste/132/base    -> origin/gh/aorenste/132/base
2025-12-04T08:53:09.5309200Z  * [new branch]                gh/aorenste/132/head    -> origin/gh/aorenste/132/head
2025-12-04T08:53:09.5309386Z  * [new branch]                gh/aorenste/134/base    -> origin/gh/aorenste/134/base
2025-12-04T08:53:09.5309571Z  * [new branch]                gh/aorenste/134/head    -> origin/gh/aorenste/134/head
2025-12-04T08:53:09.5309755Z  * [new branch]                gh/aorenste/134/orig    -> origin/gh/aorenste/134/orig
2025-12-04T08:53:09.5309942Z  * [new branch]                gh/aorenste/139/base    -> origin/gh/aorenste/139/base
2025-12-04T08:53:09.5310126Z  * [new branch]                gh/aorenste/139/head    -> origin/gh/aorenste/139/head
2025-12-04T08:53:09.5310311Z  * [new branch]                gh/aorenste/139/orig    -> origin/gh/aorenste/139/orig
2025-12-04T08:53:09.5310519Z  * [new branch]                gh/aorenste/141/base    -> origin/gh/aorenste/141/base
2025-12-04T08:53:09.5310700Z  * [new branch]                gh/aorenste/141/head    -> origin/gh/aorenste/141/head
2025-12-04T08:53:09.5310889Z  * [new branch]                gh/aorenste/145/base    -> origin/gh/aorenste/145/base
2025-12-04T08:53:09.5311072Z  * [new branch]                gh/aorenste/145/head    -> origin/gh/aorenste/145/head
2025-12-04T08:53:09.5311262Z  * [new branch]                gh/aorenste/145/orig    -> origin/gh/aorenste/145/orig
2025-12-04T08:53:09.5311447Z  * [new branch]                gh/aorenste/146/base    -> origin/gh/aorenste/146/base
2025-12-04T08:53:09.5311630Z  * [new branch]                gh/aorenste/146/head    -> origin/gh/aorenste/146/head
2025-12-04T08:53:09.5311816Z  * [new branch]                gh/aorenste/146/orig    -> origin/gh/aorenste/146/orig
2025-12-04T08:53:09.5312003Z  * [new branch]                gh/aorenste/147/base    -> origin/gh/aorenste/147/base
2025-12-04T08:53:09.5312186Z  * [new branch]                gh/aorenste/147/head    -> origin/gh/aorenste/147/head
2025-12-04T08:53:09.5312375Z  * [new branch]                gh/aorenste/147/orig    -> origin/gh/aorenste/147/orig
2025-12-04T08:53:09.5312620Z  * [new branch]                gh/aorenste/148/base    -> origin/gh/aorenste/148/base
2025-12-04T08:53:09.5312803Z  * [new branch]                gh/aorenste/148/head    -> origin/gh/aorenste/148/head
2025-12-04T08:53:09.5313036Z  * [new branch]                gh/aorenste/148/orig    -> origin/gh/aorenste/148/orig
2025-12-04T08:53:09.5313225Z  * [new branch]                gh/aorenste/149/base    -> origin/gh/aorenste/149/base
2025-12-04T08:53:09.5313407Z  * [new branch]                gh/aorenste/149/head    -> origin/gh/aorenste/149/head
2025-12-04T08:53:09.5313595Z  * [new branch]                gh/aorenste/149/orig    -> origin/gh/aorenste/149/orig
2025-12-04T08:53:09.5313785Z  * [new branch]                gh/aorenste/150/base    -> origin/gh/aorenste/150/base
2025-12-04T08:53:09.5313969Z  * [new branch]                gh/aorenste/150/head    -> origin/gh/aorenste/150/head
2025-12-04T08:53:09.5314153Z  * [new branch]                gh/aorenste/150/orig    -> origin/gh/aorenste/150/orig
2025-12-04T08:53:09.5314342Z  * [new branch]                gh/aorenste/151/base    -> origin/gh/aorenste/151/base
2025-12-04T08:53:09.5314525Z  * [new branch]                gh/aorenste/151/head    -> origin/gh/aorenste/151/head
2025-12-04T08:53:09.5314718Z  * [new branch]                gh/aorenste/151/orig    -> origin/gh/aorenste/151/orig
2025-12-04T08:53:09.5314904Z  * [new branch]                gh/aorenste/152/base    -> origin/gh/aorenste/152/base
2025-12-04T08:53:09.5315086Z  * [new branch]                gh/aorenste/152/head    -> origin/gh/aorenste/152/head
2025-12-04T08:53:09.5315272Z  * [new branch]                gh/aorenste/152/orig    -> origin/gh/aorenste/152/orig
2025-12-04T08:53:09.5315456Z  * [new branch]                gh/aorenste/153/base    -> origin/gh/aorenste/153/base
2025-12-04T08:53:09.5315637Z  * [new branch]                gh/aorenste/153/head    -> origin/gh/aorenste/153/head
2025-12-04T08:53:09.5315826Z  * [new branch]                gh/aorenste/153/orig    -> origin/gh/aorenste/153/orig
2025-12-04T08:53:09.5316007Z  * [new branch]                gh/aorenste/154/base    -> origin/gh/aorenste/154/base
2025-12-04T08:53:09.5316190Z  * [new branch]                gh/aorenste/154/head    -> origin/gh/aorenste/154/head
2025-12-04T08:53:09.5316380Z  * [new branch]                gh/aorenste/154/orig    -> origin/gh/aorenste/154/orig
2025-12-04T08:53:09.5316563Z  * [new branch]                gh/aorenste/155/base    -> origin/gh/aorenste/155/base
2025-12-04T08:53:09.5316748Z  * [new branch]                gh/aorenste/155/head    -> origin/gh/aorenste/155/head
2025-12-04T08:53:09.5316934Z  * [new branch]                gh/aorenste/155/orig    -> origin/gh/aorenste/155/orig
2025-12-04T08:53:09.5317115Z  * [new branch]                gh/aorenste/156/base    -> origin/gh/aorenste/156/base
2025-12-04T08:53:09.5317302Z  * [new branch]                gh/aorenste/156/head    -> origin/gh/aorenste/156/head
2025-12-04T08:53:09.5317486Z  * [new branch]                gh/aorenste/156/orig    -> origin/gh/aorenste/156/orig
2025-12-04T08:53:09.5317671Z  * [new branch]                gh/aorenste/157/base    -> origin/gh/aorenste/157/base
2025-12-04T08:53:09.5317859Z  * [new branch]                gh/aorenste/157/head    -> origin/gh/aorenste/157/head
2025-12-04T08:53:09.5318045Z  * [new branch]                gh/aorenste/157/orig    -> origin/gh/aorenste/157/orig
2025-12-04T08:53:09.5318227Z  * [new branch]                gh/aorenste/158/base    -> origin/gh/aorenste/158/base
2025-12-04T08:53:09.5318412Z  * [new branch]                gh/aorenste/158/head    -> origin/gh/aorenste/158/head
2025-12-04T08:53:09.5318594Z  * [new branch]                gh/aorenste/158/orig    -> origin/gh/aorenste/158/orig
2025-12-04T08:53:09.5318773Z  * [new branch]                gh/aorenste/159/base    -> origin/gh/aorenste/159/base
2025-12-04T08:53:09.5318955Z  * [new branch]                gh/aorenste/159/head    -> origin/gh/aorenste/159/head
2025-12-04T08:53:09.5319142Z  * [new branch]                gh/aorenste/159/orig    -> origin/gh/aorenste/159/orig
2025-12-04T08:53:09.5319372Z  * [new branch]                gh/avikchaudhuri/1/base -> origin/gh/avikchaudhuri/1/base
2025-12-04T08:53:09.5319599Z  * [new branch]                gh/avikchaudhuri/1/head -> origin/gh/avikchaudhuri/1/head
2025-12-04T08:53:09.5319795Z  * [new branch]                gh/avikchaudhuri/2/base -> origin/gh/avikchaudhuri/2/base
2025-12-04T08:53:09.5319992Z  * [new branch]                gh/avikchaudhuri/2/head -> origin/gh/avikchaudhuri/2/head
2025-12-04T08:53:09.5320186Z  * [new branch]                gh/avikchaudhuri/2/orig -> origin/gh/avikchaudhuri/2/orig
2025-12-04T08:53:09.5320370Z  * [new branch]                gh/bdhirsh/666/base     -> origin/gh/bdhirsh/666/base
2025-12-04T08:53:09.5320594Z  * [new branch]                gh/bdhirsh/666/head     -> origin/gh/bdhirsh/666/head
2025-12-04T08:53:09.5320776Z  * [new branch]                gh/bdhirsh/666/orig     -> origin/gh/bdhirsh/666/orig
2025-12-04T08:53:09.5320957Z  * [new branch]                gh/bdhirsh/668/base     -> origin/gh/bdhirsh/668/base
2025-12-04T08:53:09.5321139Z  * [new branch]                gh/bdhirsh/668/head     -> origin/gh/bdhirsh/668/head
2025-12-04T08:53:09.5321319Z  * [new branch]                gh/bdhirsh/668/orig     -> origin/gh/bdhirsh/668/orig
2025-12-04T08:53:09.5321496Z  * [new branch]                gh/bdhirsh/669/base     -> origin/gh/bdhirsh/669/base
2025-12-04T08:53:09.5321675Z  * [new branch]                gh/bdhirsh/669/head     -> origin/gh/bdhirsh/669/head
2025-12-04T08:53:09.5321856Z  * [new branch]                gh/bdhirsh/669/orig     -> origin/gh/bdhirsh/669/orig
2025-12-04T08:53:09.5322037Z  * [new branch]                gh/bdhirsh/670/base     -> origin/gh/bdhirsh/670/base
2025-12-04T08:53:09.5322214Z  * [new branch]                gh/bdhirsh/670/head     -> origin/gh/bdhirsh/670/head
2025-12-04T08:53:09.5322391Z  * [new branch]                gh/bdhirsh/670/orig     -> origin/gh/bdhirsh/670/orig
2025-12-04T08:53:09.5322571Z  * [new branch]                gh/bdhirsh/672/base     -> origin/gh/bdhirsh/672/base
2025-12-04T08:53:09.5322756Z  * [new branch]                gh/bdhirsh/672/head     -> origin/gh/bdhirsh/672/head
2025-12-04T08:53:09.5322938Z  * [new branch]                gh/bdhirsh/672/orig     -> origin/gh/bdhirsh/672/orig
2025-12-04T08:53:09.5323116Z  * [new branch]                gh/bdhirsh/675/base     -> origin/gh/bdhirsh/675/base
2025-12-04T08:53:09.5323297Z  * [new branch]                gh/bdhirsh/675/head     -> origin/gh/bdhirsh/675/head
2025-12-04T08:53:09.5323477Z  * [new branch]                gh/bdhirsh/675/orig     -> origin/gh/bdhirsh/675/orig
2025-12-04T08:53:09.5323662Z  * [new branch]                gh/bdhirsh/676/base     -> origin/gh/bdhirsh/676/base
2025-12-04T08:53:09.5323843Z  * [new branch]                gh/bdhirsh/676/head     -> origin/gh/bdhirsh/676/head
2025-12-04T08:53:09.5324022Z  * [new branch]                gh/bdhirsh/676/orig     -> origin/gh/bdhirsh/676/orig
2025-12-04T08:53:09.5324204Z  * [new branch]                gh/bdhirsh/677/base     -> origin/gh/bdhirsh/677/base
2025-12-04T08:53:09.5324277Z  * [new branch]                gh/bdhirsh/677/head     -> origin/gh/bdhirsh/677/head
2025-12-04T08:53:09.5324357Z  * [new branch]                gh/bdhirsh/677/orig     -> origin/gh/bdhirsh/677/orig
2025-12-04T08:53:09.5324429Z  * [new branch]                gh/bdhirsh/678/base     -> origin/gh/bdhirsh/678/base
2025-12-04T08:53:09.5324501Z  * [new branch]                gh/bdhirsh/678/head     -> origin/gh/bdhirsh/678/head
2025-12-04T08:53:09.5324571Z  * [new branch]                gh/bdhirsh/678/orig     -> origin/gh/bdhirsh/678/orig
2025-12-04T08:53:09.5324641Z  * [new branch]                gh/bdhirsh/679/base     -> origin/gh/bdhirsh/679/base
2025-12-04T08:53:09.5324713Z  * [new branch]                gh/bdhirsh/679/head     -> origin/gh/bdhirsh/679/head
2025-12-04T08:53:09.5324784Z  * [new branch]                gh/bdhirsh/679/orig     -> origin/gh/bdhirsh/679/orig
2025-12-04T08:53:09.5324854Z  * [new branch]                gh/bdhirsh/680/base     -> origin/gh/bdhirsh/680/base
2025-12-04T08:53:09.5324976Z  * [new branch]                gh/bdhirsh/680/head     -> origin/gh/bdhirsh/680/head
2025-12-04T08:53:09.5325091Z  * [new branch]                gh/bdhirsh/680/orig     -> origin/gh/bdhirsh/680/orig
2025-12-04T08:53:09.5325161Z  * [new branch]                gh/bdhirsh/681/base     -> origin/gh/bdhirsh/681/base
2025-12-04T08:53:09.5325234Z  * [new branch]                gh/bdhirsh/681/head     -> origin/gh/bdhirsh/681/head
2025-12-04T08:53:09.5325304Z  * [new branch]                gh/bdhirsh/681/orig     -> origin/gh/bdhirsh/681/orig
2025-12-04T08:53:09.5325398Z  * [new branch]                gh/benjaminglass1/101/base -> origin/gh/benjaminglass1/101/base
2025-12-04T08:53:09.5325491Z  * [new branch]                gh/benjaminglass1/101/head -> origin/gh/benjaminglass1/101/head
2025-12-04T08:53:09.5325579Z  * [new branch]                gh/benjaminglass1/101/orig -> origin/gh/benjaminglass1/101/orig
2025-12-04T08:53:09.5325671Z  * [new branch]                gh/benjaminglass1/102/base -> origin/gh/benjaminglass1/102/base
2025-12-04T08:53:09.5325765Z  * [new branch]                gh/benjaminglass1/102/head -> origin/gh/benjaminglass1/102/head
2025-12-04T08:53:09.5325852Z  * [new branch]                gh/benjaminglass1/102/orig -> origin/gh/benjaminglass1/102/orig
2025-12-04T08:53:09.5325937Z  * [new branch]                gh/benjaminglass1/106/base -> origin/gh/benjaminglass1/106/base
2025-12-04T08:53:09.5326025Z  * [new branch]                gh/benjaminglass1/106/head -> origin/gh/benjaminglass1/106/head
2025-12-04T08:53:09.5326110Z  * [new branch]                gh/benjaminglass1/106/orig -> origin/gh/benjaminglass1/106/orig
2025-12-04T08:53:09.5326198Z  * [new branch]                gh/benjaminglass1/107/base -> origin/gh/benjaminglass1/107/base
2025-12-04T08:53:09.5326283Z  * [new branch]                gh/benjaminglass1/107/head -> origin/gh/benjaminglass1/107/head
2025-12-04T08:53:09.5326368Z  * [new branch]                gh/benjaminglass1/107/orig -> origin/gh/benjaminglass1/107/orig
2025-12-04T08:53:09.5326462Z  * [new branch]                gh/benjaminglass1/108/base -> origin/gh/benjaminglass1/108/base
2025-12-04T08:53:09.5326550Z  * [new branch]                gh/benjaminglass1/108/head -> origin/gh/benjaminglass1/108/head
2025-12-04T08:53:09.5326637Z  * [new branch]                gh/benjaminglass1/108/orig -> origin/gh/benjaminglass1/108/orig
2025-12-04T08:53:09.5326724Z  * [new branch]                gh/benjaminglass1/109/base -> origin/gh/benjaminglass1/109/base
2025-12-04T08:53:09.5326808Z  * [new branch]                gh/benjaminglass1/109/head -> origin/gh/benjaminglass1/109/head
2025-12-04T08:53:09.5326893Z  * [new branch]                gh/benjaminglass1/109/orig -> origin/gh/benjaminglass1/109/orig
2025-12-04T08:53:09.5326981Z  * [new branch]                gh/benjaminglass1/97/base -> origin/gh/benjaminglass1/97/base
2025-12-04T08:53:09.5327065Z  * [new branch]                gh/benjaminglass1/97/head -> origin/gh/benjaminglass1/97/head
2025-12-04T08:53:09.5327149Z  * [new branch]                gh/benjaminglass1/97/orig -> origin/gh/benjaminglass1/97/orig
2025-12-04T08:53:09.5327232Z  * [new branch]                gh/bobrenjc93/570/base  -> origin/gh/bobrenjc93/570/base
2025-12-04T08:53:09.5327309Z  * [new branch]                gh/bobrenjc93/570/head  -> origin/gh/bobrenjc93/570/head
2025-12-04T08:53:09.5327389Z  * [new branch]                gh/bobrenjc93/570/orig  -> origin/gh/bobrenjc93/570/orig
2025-12-04T08:53:09.5327468Z  * [new branch]                gh/bobrenjc93/604/base  -> origin/gh/bobrenjc93/604/base
2025-12-04T08:53:09.5327542Z  * [new branch]                gh/bobrenjc93/604/head  -> origin/gh/bobrenjc93/604/head
2025-12-04T08:53:09.5327615Z  * [new branch]                gh/bobrenjc93/604/orig  -> origin/gh/bobrenjc93/604/orig
2025-12-04T08:53:09.5327692Z  * [new branch]                gh/bobrenjc93/638/base  -> origin/gh/bobrenjc93/638/base
2025-12-04T08:53:09.5327766Z  * [new branch]                gh/bobrenjc93/638/head  -> origin/gh/bobrenjc93/638/head
2025-12-04T08:53:09.5327873Z  * [new branch]                gh/bobrenjc93/638/orig  -> origin/gh/bobrenjc93/638/orig
2025-12-04T08:53:09.5327982Z  * [new branch]                gh/bobrenjc93/653/base  -> origin/gh/bobrenjc93/653/base
2025-12-04T08:53:09.5328057Z  * [new branch]                gh/bobrenjc93/653/head  -> origin/gh/bobrenjc93/653/head
2025-12-04T08:53:09.5328135Z  * [new branch]                gh/bobrenjc93/653/orig  -> origin/gh/bobrenjc93/653/orig
2025-12-04T08:53:09.5328210Z  * [new branch]                gh/bobrenjc93/654/base  -> origin/gh/bobrenjc93/654/base
2025-12-04T08:53:09.5328283Z  * [new branch]                gh/bobrenjc93/654/head  -> origin/gh/bobrenjc93/654/head
2025-12-04T08:53:09.5328359Z  * [new branch]                gh/bobrenjc93/654/orig  -> origin/gh/bobrenjc93/654/orig
2025-12-04T08:53:09.5328433Z  * [new branch]                gh/bobrenjc93/657/base  -> origin/gh/bobrenjc93/657/base
2025-12-04T08:53:09.5328509Z  * [new branch]                gh/bobrenjc93/657/head  -> origin/gh/bobrenjc93/657/head
2025-12-04T08:53:09.5328586Z  * [new branch]                gh/bobrenjc93/657/orig  -> origin/gh/bobrenjc93/657/orig
2025-12-04T08:53:09.5328663Z  * [new branch]                gh/bobrenjc93/672/base  -> origin/gh/bobrenjc93/672/base
2025-12-04T08:53:09.5328738Z  * [new branch]                gh/bobrenjc93/672/head  -> origin/gh/bobrenjc93/672/head
2025-12-04T08:53:09.5328817Z  * [new branch]                gh/bobrenjc93/672/orig  -> origin/gh/bobrenjc93/672/orig
2025-12-04T08:53:09.5328892Z  * [new branch]                gh/bobrenjc93/679/base  -> origin/gh/bobrenjc93/679/base
2025-12-04T08:53:09.5328968Z  * [new branch]                gh/bobrenjc93/679/head  -> origin/gh/bobrenjc93/679/head
2025-12-04T08:53:09.5329045Z  * [new branch]                gh/bobrenjc93/679/orig  -> origin/gh/bobrenjc93/679/orig
2025-12-04T08:53:09.5329119Z  * [new branch]                gh/bobrenjc93/680/base  -> origin/gh/bobrenjc93/680/base
2025-12-04T08:53:09.5329195Z  * [new branch]                gh/bobrenjc93/680/head  -> origin/gh/bobrenjc93/680/head
2025-12-04T08:53:09.5329271Z  * [new branch]                gh/bobrenjc93/680/orig  -> origin/gh/bobrenjc93/680/orig
2025-12-04T08:53:09.5329345Z  * [new branch]                gh/bobrenjc93/681/base  -> origin/gh/bobrenjc93/681/base
2025-12-04T08:53:09.5329420Z  * [new branch]                gh/bobrenjc93/681/head  -> origin/gh/bobrenjc93/681/head
2025-12-04T08:53:09.5329494Z  * [new branch]                gh/bobrenjc93/681/orig  -> origin/gh/bobrenjc93/681/orig
2025-12-04T08:53:09.5329568Z  * [new branch]                gh/bobrenjc93/682/base  -> origin/gh/bobrenjc93/682/base
2025-12-04T08:53:09.5329644Z  * [new branch]                gh/bobrenjc93/682/head  -> origin/gh/bobrenjc93/682/head
2025-12-04T08:53:09.5329720Z  * [new branch]                gh/bobrenjc93/682/orig  -> origin/gh/bobrenjc93/682/orig
2025-12-04T08:53:09.5329794Z  * [new branch]                gh/bobrenjc93/683/base  -> origin/gh/bobrenjc93/683/base
2025-12-04T08:53:09.5329872Z  * [new branch]                gh/bobrenjc93/683/head  -> origin/gh/bobrenjc93/683/head
2025-12-04T08:53:09.5329946Z  * [new branch]                gh/bobrenjc93/683/orig  -> origin/gh/bobrenjc93/683/orig
2025-12-04T08:53:09.5330019Z  * [new branch]                gh/bobrenjc93/684/base  -> origin/gh/bobrenjc93/684/base
2025-12-04T08:53:09.5330098Z  * [new branch]                gh/bobrenjc93/684/head  -> origin/gh/bobrenjc93/684/head
2025-12-04T08:53:09.5330174Z  * [new branch]                gh/bobrenjc93/684/orig  -> origin/gh/bobrenjc93/684/orig
2025-12-04T08:53:09.5330249Z  * [new branch]                gh/bobrenjc93/685/base  -> origin/gh/bobrenjc93/685/base
2025-12-04T08:53:09.5330329Z  * [new branch]                gh/bobrenjc93/685/head  -> origin/gh/bobrenjc93/685/head
2025-12-04T08:53:09.5330426Z  * [new branch]                gh/bobrenjc93/685/orig  -> origin/gh/bobrenjc93/685/orig
2025-12-04T08:53:09.5330502Z  * [new branch]                gh/bobrenjc93/686/base  -> origin/gh/bobrenjc93/686/base
2025-12-04T08:53:09.5330628Z  * [new branch]                gh/bobrenjc93/686/head  -> origin/gh/bobrenjc93/686/head
2025-12-04T08:53:09.5330746Z  * [new branch]                gh/bobrenjc93/686/orig  -> origin/gh/bobrenjc93/686/orig
2025-12-04T08:53:09.5330820Z  * [new branch]                gh/bobrenjc93/687/base  -> origin/gh/bobrenjc93/687/base
2025-12-04T08:53:09.5330899Z  * [new branch]                gh/bobrenjc93/687/head  -> origin/gh/bobrenjc93/687/head
2025-12-04T08:53:09.5330974Z  * [new branch]                gh/bobrenjc93/687/orig  -> origin/gh/bobrenjc93/687/orig
2025-12-04T08:53:09.5331051Z  * [new branch]                gh/bobrenjc93/688/base  -> origin/gh/bobrenjc93/688/base
2025-12-04T08:53:09.5331125Z  * [new branch]                gh/bobrenjc93/688/head  -> origin/gh/bobrenjc93/688/head
2025-12-04T08:53:09.5331199Z  * [new branch]                gh/bobrenjc93/688/orig  -> origin/gh/bobrenjc93/688/orig
2025-12-04T08:53:09.5331279Z  * [new branch]                gh/bobrenjc93/689/base  -> origin/gh/bobrenjc93/689/base
2025-12-04T08:53:09.5331355Z  * [new branch]                gh/bobrenjc93/689/head  -> origin/gh/bobrenjc93/689/head
2025-12-04T08:53:09.5331431Z  * [new branch]                gh/bobrenjc93/689/orig  -> origin/gh/bobrenjc93/689/orig
2025-12-04T08:53:09.5331507Z  * [new branch]                gh/bobrenjc93/690/base  -> origin/gh/bobrenjc93/690/base
2025-12-04T08:53:09.5331580Z  * [new branch]                gh/bobrenjc93/690/head  -> origin/gh/bobrenjc93/690/head
2025-12-04T08:53:09.5331654Z  * [new branch]                gh/bobrenjc93/690/orig  -> origin/gh/bobrenjc93/690/orig
2025-12-04T08:53:09.5331730Z  * [new branch]                gh/bobrenjc93/691/base  -> origin/gh/bobrenjc93/691/base
2025-12-04T08:53:09.5331803Z  * [new branch]                gh/bobrenjc93/691/head  -> origin/gh/bobrenjc93/691/head
2025-12-04T08:53:09.5331878Z  * [new branch]                gh/bobrenjc93/691/orig  -> origin/gh/bobrenjc93/691/orig
2025-12-04T08:53:09.5331956Z  * [new branch]                gh/bobrenjc93/692/base  -> origin/gh/bobrenjc93/692/base
2025-12-04T08:53:09.5332033Z  * [new branch]                gh/bobrenjc93/692/head  -> origin/gh/bobrenjc93/692/head
2025-12-04T08:53:09.5332108Z  * [new branch]                gh/bobrenjc93/692/orig  -> origin/gh/bobrenjc93/692/orig
2025-12-04T08:53:09.5332185Z  * [new branch]                gh/bobrenjc93/693/base  -> origin/gh/bobrenjc93/693/base
2025-12-04T08:53:09.5332259Z  * [new branch]                gh/bobrenjc93/693/head  -> origin/gh/bobrenjc93/693/head
2025-12-04T08:53:09.5332332Z  * [new branch]                gh/bobrenjc93/693/orig  -> origin/gh/bobrenjc93/693/orig
2025-12-04T08:53:09.5332409Z  * [new branch]                gh/bobrenjc93/694/base  -> origin/gh/bobrenjc93/694/base
2025-12-04T08:53:09.5332482Z  * [new branch]                gh/bobrenjc93/694/head  -> origin/gh/bobrenjc93/694/head
2025-12-04T08:53:09.5332556Z  * [new branch]                gh/bobrenjc93/694/orig  -> origin/gh/bobrenjc93/694/orig
2025-12-04T08:53:09.5332634Z  * [new branch]                gh/bobrenjc93/695/base  -> origin/gh/bobrenjc93/695/base
2025-12-04T08:53:09.5332709Z  * [new branch]                gh/bobrenjc93/695/head  -> origin/gh/bobrenjc93/695/head
2025-12-04T08:53:09.5332788Z  * [new branch]                gh/bobrenjc93/695/orig  -> origin/gh/bobrenjc93/695/orig
2025-12-04T08:53:09.5332856Z  * [new branch]                gh/c00w/23/base         -> origin/gh/c00w/23/base
2025-12-04T08:53:09.5332925Z  * [new branch]                gh/c00w/23/head         -> origin/gh/c00w/23/head
2025-12-04T08:53:09.5332991Z  * [new branch]                gh/c00w/53/base         -> origin/gh/c00w/53/base
2025-12-04T08:53:09.5333055Z  * [new branch]                gh/c00w/53/head         -> origin/gh/c00w/53/head
2025-12-04T08:53:09.5333120Z  * [new branch]                gh/c00w/53/orig         -> origin/gh/c00w/53/orig
2025-12-04T08:53:09.5333187Z  * [new branch]                gh/c00w/54/base         -> origin/gh/c00w/54/base
2025-12-04T08:53:09.5333281Z  * [new branch]                gh/c00w/54/head         -> origin/gh/c00w/54/head
2025-12-04T08:53:09.5333373Z  * [new branch]                gh/c00w/54/orig         -> origin/gh/c00w/54/orig
2025-12-04T08:53:09.5333443Z  * [new branch]                gh/c00w/56/base         -> origin/gh/c00w/56/base
2025-12-04T08:53:09.5333510Z  * [new branch]                gh/c00w/56/head         -> origin/gh/c00w/56/head
2025-12-04T08:53:09.5333574Z  * [new branch]                gh/c00w/56/orig         -> origin/gh/c00w/56/orig
2025-12-04T08:53:09.5333640Z  * [new branch]                gh/c00w/57/base         -> origin/gh/c00w/57/base
2025-12-04T08:53:09.5333703Z  * [new branch]                gh/c00w/57/head         -> origin/gh/c00w/57/head
2025-12-04T08:53:09.5333766Z  * [new branch]                gh/c00w/57/orig         -> origin/gh/c00w/57/orig
2025-12-04T08:53:09.5333834Z  * [new branch]                gh/c00w/58/base         -> origin/gh/c00w/58/base
2025-12-04T08:53:09.5333900Z  * [new branch]                gh/c00w/58/head         -> origin/gh/c00w/58/head
2025-12-04T08:53:09.5333963Z  * [new branch]                gh/c00w/58/orig         -> origin/gh/c00w/58/orig
2025-12-04T08:53:09.5334044Z  * [new branch]                gh/clee2000/1/base      -> origin/gh/clee2000/1/base
2025-12-04T08:53:09.5334117Z  * [new branch]                gh/clee2000/1/head      -> origin/gh/clee2000/1/head
2025-12-04T08:53:09.5334187Z  * [new branch]                gh/clee2000/1/orig      -> origin/gh/clee2000/1/orig
2025-12-04T08:53:09.5334270Z  * [new branch]                gh/coconutruben/1/base  -> origin/gh/coconutruben/1/base
2025-12-04T08:53:09.5334347Z  * [new branch]                gh/coconutruben/1/head  -> origin/gh/coconutruben/1/head
2025-12-04T08:53:09.5334428Z  * [new branch]                gh/coconutruben/55/base -> origin/gh/coconutruben/55/base
2025-12-04T08:53:09.5334507Z  * [new branch]                gh/coconutruben/55/head -> origin/gh/coconutruben/55/head
2025-12-04T08:53:09.5334586Z  * [new branch]                gh/coconutruben/55/orig -> origin/gh/coconutruben/55/orig
2025-12-04T08:53:09.5334671Z  * [new branch]                gh/coconutruben/57/base -> origin/gh/coconutruben/57/base
2025-12-04T08:53:09.5334751Z  * [new branch]                gh/coconutruben/57/head -> origin/gh/coconutruben/57/head
2025-12-04T08:53:09.5334828Z  * [new branch]                gh/coconutruben/57/orig -> origin/gh/coconutruben/57/orig
2025-12-04T08:53:09.5334909Z  * [new branch]                gh/coconutruben/70/base -> origin/gh/coconutruben/70/base
2025-12-04T08:53:09.5334985Z  * [new branch]                gh/coconutruben/70/head -> origin/gh/coconutruben/70/head
2025-12-04T08:53:09.5335063Z  * [new branch]                gh/coconutruben/70/orig -> origin/gh/coconutruben/70/orig
2025-12-04T08:53:09.5335142Z  * [new branch]                gh/coconutruben/71/base -> origin/gh/coconutruben/71/base
2025-12-04T08:53:09.5335219Z  * [new branch]                gh/coconutruben/71/head -> origin/gh/coconutruben/71/head
2025-12-04T08:53:09.5335299Z  * [new branch]                gh/coconutruben/71/orig -> origin/gh/coconutruben/71/orig
2025-12-04T08:53:09.5335383Z  * [new branch]                gh/coconutruben/72/base -> origin/gh/coconutruben/72/base
2025-12-04T08:53:09.5335461Z  * [new branch]                gh/coconutruben/72/head -> origin/gh/coconutruben/72/head
2025-12-04T08:53:09.5335538Z  * [new branch]                gh/coconutruben/72/orig -> origin/gh/coconutruben/72/orig
2025-12-04T08:53:09.5335622Z  * [new branch]                gh/coconutruben/73/base -> origin/gh/coconutruben/73/base
2025-12-04T08:53:09.5335700Z  * [new branch]                gh/coconutruben/73/head -> origin/gh/coconutruben/73/head
2025-12-04T08:53:09.5335777Z  * [new branch]                gh/coconutruben/73/orig -> origin/gh/coconutruben/73/orig
2025-12-04T08:53:09.5335857Z  * [new branch]                gh/coconutruben/74/base -> origin/gh/coconutruben/74/base
2025-12-04T08:53:09.5335934Z  * [new branch]                gh/coconutruben/74/head -> origin/gh/coconutruben/74/head
2025-12-04T08:53:09.5336047Z  * [new branch]                gh/coconutruben/74/orig -> origin/gh/coconutruben/74/orig
2025-12-04T08:53:09.5336155Z  * [new branch]                gh/coconutruben/79/base -> origin/gh/coconutruben/79/base
2025-12-04T08:53:09.5336232Z  * [new branch]                gh/coconutruben/79/head -> origin/gh/coconutruben/79/head
2025-12-04T08:53:09.5336311Z  * [new branch]                gh/coconutruben/79/orig -> origin/gh/coconutruben/79/orig
2025-12-04T08:53:09.5336387Z  * [new branch]                gh/coconutruben/80/base -> origin/gh/coconutruben/80/base
2025-12-04T08:53:09.5336463Z  * [new branch]                gh/coconutruben/80/head -> origin/gh/coconutruben/80/head
2025-12-04T08:53:09.5336542Z  * [new branch]                gh/coconutruben/80/orig -> origin/gh/coconutruben/80/orig
2025-12-04T08:53:09.5336618Z  * [new branch]                gh/coconutruben/82/base -> origin/gh/coconutruben/82/base
2025-12-04T08:53:09.5336695Z  * [new branch]                gh/coconutruben/82/head -> origin/gh/coconutruben/82/head
2025-12-04T08:53:09.5336776Z  * [new branch]                gh/coconutruben/82/orig -> origin/gh/coconutruben/82/orig
2025-12-04T08:53:09.5336855Z  * [new branch]                gh/coconutruben/83/base -> origin/gh/coconutruben/83/base
2025-12-04T08:53:09.5336933Z  * [new branch]                gh/coconutruben/83/head -> origin/gh/coconutruben/83/head
2025-12-04T08:53:09.5337011Z  * [new branch]                gh/coconutruben/83/orig -> origin/gh/coconutruben/83/orig
2025-12-04T08:53:09.5337087Z  * [new branch]                gh/coconutruben/84/base -> origin/gh/coconutruben/84/base
2025-12-04T08:53:09.5337163Z  * [new branch]                gh/coconutruben/84/head -> origin/gh/coconutruben/84/head
2025-12-04T08:53:09.5337242Z  * [new branch]                gh/coconutruben/84/orig -> origin/gh/coconutruben/84/orig
2025-12-04T08:53:09.5337318Z  * [new branch]                gh/coconutruben/85/base -> origin/gh/coconutruben/85/base
2025-12-04T08:53:09.5337397Z  * [new branch]                gh/coconutruben/85/head -> origin/gh/coconutruben/85/head
2025-12-04T08:53:09.5337481Z  * [new branch]                gh/coconutruben/85/orig -> origin/gh/coconutruben/85/orig
2025-12-04T08:53:09.5337557Z  * [new branch]                gh/coconutruben/86/base -> origin/gh/coconutruben/86/base
2025-12-04T08:53:09.5337635Z  * [new branch]                gh/coconutruben/86/head -> origin/gh/coconutruben/86/head
2025-12-04T08:53:09.5337715Z  * [new branch]                gh/coconutruben/86/orig -> origin/gh/coconutruben/86/orig
2025-12-04T08:53:09.5337794Z  * [new branch]                gh/colinchan15/1/base   -> origin/gh/colinchan15/1/base
2025-12-04T08:53:09.5337873Z  * [new branch]                gh/colinchan15/1/head   -> origin/gh/colinchan15/1/head
2025-12-04T08:53:09.5337948Z  * [new branch]                gh/colinchan15/2/base   -> origin/gh/colinchan15/2/base
2025-12-04T08:53:09.5338023Z  * [new branch]                gh/colinchan15/2/head   -> origin/gh/colinchan15/2/head
2025-12-04T08:53:09.5338104Z  * [new branch]                gh/colinchan15/3/base   -> origin/gh/colinchan15/3/base
2025-12-04T08:53:09.5338183Z  * [new branch]                gh/colinchan15/3/head   -> origin/gh/colinchan15/3/head
2025-12-04T08:53:09.5338258Z  * [new branch]                gh/colinchan15/6/base   -> origin/gh/colinchan15/6/base
2025-12-04T08:53:09.5338333Z  * [new branch]                gh/colinchan15/6/head   -> origin/gh/colinchan15/6/head
2025-12-04T08:53:09.5338400Z  * [new branch]                gh/d4l3k/1/base         -> origin/gh/d4l3k/1/base
2025-12-04T08:53:09.5338464Z  * [new branch]                gh/d4l3k/1/head         -> origin/gh/d4l3k/1/head
2025-12-04T08:53:09.5338530Z  * [new branch]                gh/d4l3k/2/base         -> origin/gh/d4l3k/2/base
2025-12-04T08:53:09.5338595Z  * [new branch]                gh/d4l3k/2/head         -> origin/gh/d4l3k/2/head
2025-12-04T08:53:09.5338659Z  * [new branch]                gh/d4l3k/2/orig         -> origin/gh/d4l3k/2/orig
2025-12-04T08:53:09.5338759Z  * [new branch]                gh/d4l3k/3/base         -> origin/gh/d4l3k/3/base
2025-12-04T08:53:09.5338866Z  * [new branch]                gh/d4l3k/3/head         -> origin/gh/d4l3k/3/head
2025-12-04T08:53:09.5338929Z  * [new branch]                gh/d4l3k/3/orig         -> origin/gh/d4l3k/3/orig
2025-12-04T08:53:09.5338996Z  * [new branch]                gh/d4l3k/4/base         -> origin/gh/d4l3k/4/base
2025-12-04T08:53:09.5339059Z  * [new branch]                gh/d4l3k/4/head         -> origin/gh/d4l3k/4/head
2025-12-04T08:53:09.5339122Z  * [new branch]                gh/d4l3k/4/orig         -> origin/gh/d4l3k/4/orig
2025-12-04T08:53:09.5339187Z  * [new branch]                gh/d4l3k/5/base         -> origin/gh/d4l3k/5/base
2025-12-04T08:53:09.5339251Z  * [new branch]                gh/d4l3k/5/orig         -> origin/gh/d4l3k/5/orig
2025-12-04T08:53:09.5339340Z  * [new branch]                gh/davidberard98/392/base -> origin/gh/davidberard98/392/base
2025-12-04T08:53:09.5339430Z  * [new branch]                gh/davidberard98/392/head -> origin/gh/davidberard98/392/head
2025-12-04T08:53:09.5339515Z  * [new branch]                gh/davidberard98/392/orig -> origin/gh/davidberard98/392/orig
2025-12-04T08:53:09.5339600Z  * [new branch]                gh/davidberard98/399/base -> origin/gh/davidberard98/399/base
2025-12-04T08:53:09.5339682Z  * [new branch]                gh/davidberard98/399/head -> origin/gh/davidberard98/399/head
2025-12-04T08:53:09.5339764Z  * [new branch]                gh/davidberard98/399/orig -> origin/gh/davidberard98/399/orig
2025-12-04T08:53:09.5339842Z  * [new branch]                gh/desertfire/605/base  -> origin/gh/desertfire/605/base
2025-12-04T08:53:09.5339918Z  * [new branch]                gh/desertfire/605/head  -> origin/gh/desertfire/605/head
2025-12-04T08:53:09.5339995Z  * [new branch]                gh/desertfire/605/orig  -> origin/gh/desertfire/605/orig
2025-12-04T08:53:09.5340072Z  * [new branch]                gh/desertfire/606/base  -> origin/gh/desertfire/606/base
2025-12-04T08:53:09.5340148Z  * [new branch]                gh/desertfire/606/head  -> origin/gh/desertfire/606/head
2025-12-04T08:53:09.5340225Z  * [new branch]                gh/desertfire/606/orig  -> origin/gh/desertfire/606/orig
2025-12-04T08:53:09.5340305Z  * [new branch]                gh/desertfire/607/base  -> origin/gh/desertfire/607/base
2025-12-04T08:53:09.5340379Z  * [new branch]                gh/desertfire/607/head  -> origin/gh/desertfire/607/head
2025-12-04T08:53:09.5340491Z  * [new branch]                gh/desertfire/607/orig  -> origin/gh/desertfire/607/orig
2025-12-04T08:53:09.5340566Z  * [new branch]                gh/desertfire/608/base  -> origin/gh/desertfire/608/base
2025-12-04T08:53:09.5340640Z  * [new branch]                gh/desertfire/608/head  -> origin/gh/desertfire/608/head
2025-12-04T08:53:09.5340714Z  * [new branch]                gh/desertfire/608/orig  -> origin/gh/desertfire/608/orig
2025-12-04T08:53:09.5340791Z  * [new branch]                gh/desertfire/609/base  -> origin/gh/desertfire/609/base
2025-12-04T08:53:09.5340865Z  * [new branch]                gh/desertfire/609/head  -> origin/gh/desertfire/609/head
2025-12-04T08:53:09.5340940Z  * [new branch]                gh/desertfire/609/orig  -> origin/gh/desertfire/609/orig
2025-12-04T08:53:09.5341018Z  * [new branch]                gh/desertfire/610/base  -> origin/gh/desertfire/610/base
2025-12-04T08:53:09.5341094Z  * [new branch]                gh/desertfire/610/head  -> origin/gh/desertfire/610/head
2025-12-04T08:53:09.5341169Z  * [new branch]                gh/desertfire/610/orig  -> origin/gh/desertfire/610/orig
2025-12-04T08:53:09.5341242Z  * [new branch]                gh/desertfire/611/base  -> origin/gh/desertfire/611/base
2025-12-04T08:53:09.5341315Z  * [new branch]                gh/desertfire/611/head  -> origin/gh/desertfire/611/head
2025-12-04T08:53:09.5341393Z  * [new branch]                gh/desertfire/611/orig  -> origin/gh/desertfire/611/orig
2025-12-04T08:53:09.5341508Z  * [new branch]                gh/desertfire/612/base  -> origin/gh/desertfire/612/base
2025-12-04T08:53:09.5341625Z  * [new branch]                gh/desertfire/612/head  -> origin/gh/desertfire/612/head
2025-12-04T08:53:09.5341701Z  * [new branch]                gh/desertfire/612/orig  -> origin/gh/desertfire/612/orig
2025-12-04T08:53:09.5341774Z  * [new branch]                gh/desertfire/613/base  -> origin/gh/desertfire/613/base
2025-12-04T08:53:09.5341849Z  * [new branch]                gh/desertfire/613/head  -> origin/gh/desertfire/613/head
2025-12-04T08:53:09.5341926Z  * [new branch]                gh/desertfire/613/orig  -> origin/gh/desertfire/613/orig
2025-12-04T08:53:09.5341998Z  * [new branch]                gh/desertfire/614/base  -> origin/gh/desertfire/614/base
2025-12-04T08:53:09.5342071Z  * [new branch]                gh/desertfire/614/head  -> origin/gh/desertfire/614/head
2025-12-04T08:53:09.5342146Z  * [new branch]                gh/desertfire/614/orig  -> origin/gh/desertfire/614/orig
2025-12-04T08:53:09.5342223Z  * [new branch]                gh/desertfire/615/base  -> origin/gh/desertfire/615/base
2025-12-04T08:53:09.5342297Z  * [new branch]                gh/desertfire/615/head  -> origin/gh/desertfire/615/head
2025-12-04T08:53:09.5342374Z  * [new branch]                gh/desertfire/615/orig  -> origin/gh/desertfire/615/orig
2025-12-04T08:53:09.5342447Z  * [new branch]                gh/desertfire/616/base  -> origin/gh/desertfire/616/base
2025-12-04T08:53:09.5342520Z  * [new branch]                gh/desertfire/616/head  -> origin/gh/desertfire/616/head
2025-12-04T08:53:09.5342595Z  * [new branch]                gh/desertfire/616/orig  -> origin/gh/desertfire/616/orig
2025-12-04T08:53:09.5342667Z  * [new branch]                gh/desertfire/617/base  -> origin/gh/desertfire/617/base
2025-12-04T08:53:09.5342740Z  * [new branch]                gh/desertfire/617/head  -> origin/gh/desertfire/617/head
2025-12-04T08:53:09.5342817Z  * [new branch]                gh/desertfire/617/orig  -> origin/gh/desertfire/617/orig
2025-12-04T08:53:09.5342890Z  * [new branch]                gh/dharakk/1/base       -> origin/gh/dharakk/1/base
2025-12-04T08:53:09.5342964Z  * [new branch]                gh/dharakk/1/head       -> origin/gh/dharakk/1/head
2025-12-04T08:53:09.5343037Z  * [new branch]                gh/drisspg/170/base     -> origin/gh/drisspg/170/base
2025-12-04T08:53:09.5343109Z  * [new branch]                gh/drisspg/170/head     -> origin/gh/drisspg/170/head
2025-12-04T08:53:09.5343181Z  * [new branch]                gh/drisspg/170/orig     -> origin/gh/drisspg/170/orig
2025-12-04T08:53:09.5343251Z  * [new branch]                gh/drisspg/182/base     -> origin/gh/drisspg/182/base
2025-12-04T08:53:09.5343322Z  * [new branch]                gh/drisspg/182/head     -> origin/gh/drisspg/182/head
2025-12-04T08:53:09.5343394Z  * [new branch]                gh/drisspg/183/base     -> origin/gh/drisspg/183/base
2025-12-04T08:53:09.5343464Z  * [new branch]                gh/drisspg/183/head     -> origin/gh/drisspg/183/head
2025-12-04T08:53:09.5343534Z  * [new branch]                gh/drisspg/184/base     -> origin/gh/drisspg/184/base
2025-12-04T08:53:09.5343606Z  * [new branch]                gh/drisspg/184/head     -> origin/gh/drisspg/184/head
2025-12-04T08:53:09.5343676Z  * [new branch]                gh/drisspg/185/base     -> origin/gh/drisspg/185/base
2025-12-04T08:53:09.5343746Z  * [new branch]                gh/drisspg/185/head     -> origin/gh/drisspg/185/head
2025-12-04T08:53:09.5343817Z  * [new branch]                gh/drisspg/194/base     -> origin/gh/drisspg/194/base
2025-12-04T08:53:09.5343887Z  * [new branch]                gh/drisspg/194/head     -> origin/gh/drisspg/194/head
2025-12-04T08:53:09.5343956Z  * [new branch]                gh/drisspg/194/orig     -> origin/gh/drisspg/194/orig
2025-12-04T08:53:09.5344028Z  * [new branch]                gh/drisspg/200/base     -> origin/gh/drisspg/200/base
2025-12-04T08:53:09.5344123Z  * [new branch]                gh/drisspg/200/head     -> origin/gh/drisspg/200/head
2025-12-04T08:53:09.5344193Z  * [new branch]                gh/drisspg/200/orig     -> origin/gh/drisspg/200/orig
2025-12-04T08:53:09.5344289Z  * [new branch]                gh/drisspg/218/base     -> origin/gh/drisspg/218/base
2025-12-04T08:53:09.5344361Z  * [new branch]                gh/drisspg/218/head     -> origin/gh/drisspg/218/head
2025-12-04T08:53:09.5344433Z  * [new branch]                gh/drisspg/218/orig     -> origin/gh/drisspg/218/orig
2025-12-04T08:53:09.5344505Z  * [new branch]                gh/drisspg/219/base     -> origin/gh/drisspg/219/base
2025-12-04T08:53:09.5344576Z  * [new branch]                gh/drisspg/219/head     -> origin/gh/drisspg/219/head
2025-12-04T08:53:09.5344652Z  * [new branch]                gh/drisspg/219/orig     -> origin/gh/drisspg/219/orig
2025-12-04T08:53:09.5344724Z  * [new branch]                gh/drisspg/220/base     -> origin/gh/drisspg/220/base
2025-12-04T08:53:09.5344796Z  * [new branch]                gh/drisspg/220/head     -> origin/gh/drisspg/220/head
2025-12-04T08:53:09.5344867Z  * [new branch]                gh/drisspg/220/orig     -> origin/gh/drisspg/220/orig
2025-12-04T08:53:09.5344936Z  * [new branch]                gh/drisspg/221/base     -> origin/gh/drisspg/221/base
2025-12-04T08:53:09.5345007Z  * [new branch]                gh/drisspg/221/head     -> origin/gh/drisspg/221/head
2025-12-04T08:53:09.5345078Z  * [new branch]                gh/drisspg/221/orig     -> origin/gh/drisspg/221/orig
2025-12-04T08:53:09.5345147Z  * [new branch]                gh/drisspg/222/base     -> origin/gh/drisspg/222/base
2025-12-04T08:53:09.5345217Z  * [new branch]                gh/drisspg/222/head     -> origin/gh/drisspg/222/head
2025-12-04T08:53:09.5345292Z  * [new branch]                gh/drisspg/222/orig     -> origin/gh/drisspg/222/orig
2025-12-04T08:53:09.5345361Z  * [new branch]                gh/drisspg/223/base     -> origin/gh/drisspg/223/base
2025-12-04T08:53:09.5345431Z  * [new branch]                gh/drisspg/223/head     -> origin/gh/drisspg/223/head
2025-12-04T08:53:09.5345502Z  * [new branch]                gh/drisspg/223/orig     -> origin/gh/drisspg/223/orig
2025-12-04T08:53:09.5345573Z  * [new branch]                gh/drisspg/224/base     -> origin/gh/drisspg/224/base
2025-12-04T08:53:09.5345642Z  * [new branch]                gh/drisspg/224/head     -> origin/gh/drisspg/224/head
2025-12-04T08:53:09.5345714Z  * [new branch]                gh/drisspg/224/orig     -> origin/gh/drisspg/224/orig
2025-12-04T08:53:09.5345783Z  * [new branch]                gh/drisspg/225/base     -> origin/gh/drisspg/225/base
2025-12-04T08:53:09.5345852Z  * [new branch]                gh/drisspg/225/head     -> origin/gh/drisspg/225/head
2025-12-04T08:53:09.5345925Z  * [new branch]                gh/drisspg/225/orig     -> origin/gh/drisspg/225/orig
2025-12-04T08:53:09.5345995Z  * [new branch]                gh/drisspg/226/base     -> origin/gh/drisspg/226/base
2025-12-04T08:53:09.5346068Z  * [new branch]                gh/drisspg/226/head     -> origin/gh/drisspg/226/head
2025-12-04T08:53:09.5346139Z  * [new branch]                gh/drisspg/226/orig     -> origin/gh/drisspg/226/orig
2025-12-04T08:53:09.5346210Z  * [new branch]                gh/drisspg/227/base     -> origin/gh/drisspg/227/base
2025-12-04T08:53:09.5346281Z  * [new branch]                gh/drisspg/227/head     -> origin/gh/drisspg/227/head
2025-12-04T08:53:09.5346350Z  * [new branch]                gh/drisspg/227/orig     -> origin/gh/drisspg/227/orig
2025-12-04T08:53:09.5346418Z  * [new branch]                gh/drisspg/228/base     -> origin/gh/drisspg/228/base
2025-12-04T08:53:09.5346490Z  * [new branch]                gh/drisspg/228/head     -> origin/gh/drisspg/228/head
2025-12-04T08:53:09.5346561Z  * [new branch]                gh/drisspg/228/orig     -> origin/gh/drisspg/228/orig
2025-12-04T08:53:09.5346633Z  * [new branch]                gh/drisspg/229/base     -> origin/gh/drisspg/229/base
2025-12-04T08:53:09.5346728Z  * [new branch]                gh/drisspg/229/head     -> origin/gh/drisspg/229/head
2025-12-04T08:53:09.5346798Z  * [new branch]                gh/drisspg/229/orig     -> origin/gh/drisspg/229/orig
2025-12-04T08:53:09.5346894Z  * [new branch]                gh/drisspg/230/base     -> origin/gh/drisspg/230/base
2025-12-04T08:53:09.5346965Z  * [new branch]                gh/drisspg/230/head     -> origin/gh/drisspg/230/head
2025-12-04T08:53:09.5347034Z  * [new branch]                gh/drisspg/230/orig     -> origin/gh/drisspg/230/orig
2025-12-04T08:53:09.5347108Z  * [new branch]                gh/dsjohns2/1/base      -> origin/gh/dsjohns2/1/base
2025-12-04T08:53:09.5347182Z  * [new branch]                gh/dsjohns2/1/head      -> origin/gh/dsjohns2/1/head
2025-12-04T08:53:09.5347262Z  * [new branch]                gh/dzmitry-huba/1/base  -> origin/gh/dzmitry-huba/1/base
2025-12-04T08:53:09.5347339Z  * [new branch]                gh/dzmitry-huba/1/head  -> origin/gh/dzmitry-huba/1/head
2025-12-04T08:53:09.5347424Z  * [new branch]                gh/dzmitry-huba/12/base -> origin/gh/dzmitry-huba/12/base
2025-12-04T08:53:09.5347503Z  * [new branch]                gh/dzmitry-huba/12/head -> origin/gh/dzmitry-huba/12/head
2025-12-04T08:53:09.5347582Z  * [new branch]                gh/dzmitry-huba/12/orig -> origin/gh/dzmitry-huba/12/orig
2025-12-04T08:53:09.5347660Z  * [new branch]                gh/dzmitry-huba/13/base -> origin/gh/dzmitry-huba/13/base
2025-12-04T08:53:09.5347734Z  * [new branch]                gh/dzmitry-huba/13/head -> origin/gh/dzmitry-huba/13/head
2025-12-04T08:53:09.5347809Z  * [new branch]                gh/dzmitry-huba/13/orig -> origin/gh/dzmitry-huba/13/orig
2025-12-04T08:53:09.5347886Z  * [new branch]                gh/dzmitry-huba/14/base -> origin/gh/dzmitry-huba/14/base
2025-12-04T08:53:09.5347960Z  * [new branch]                gh/dzmitry-huba/14/head -> origin/gh/dzmitry-huba/14/head
2025-12-04T08:53:09.5348037Z  * [new branch]                gh/dzmitry-huba/14/orig -> origin/gh/dzmitry-huba/14/orig
2025-12-04T08:53:09.5348113Z  * [new branch]                gh/dzmitry-huba/15/base -> origin/gh/dzmitry-huba/15/base
2025-12-04T08:53:09.5348189Z  * [new branch]                gh/dzmitry-huba/15/head -> origin/gh/dzmitry-huba/15/head
2025-12-04T08:53:09.5348265Z  * [new branch]                gh/dzmitry-huba/15/orig -> origin/gh/dzmitry-huba/15/orig
2025-12-04T08:53:09.5348341Z  * [new branch]                gh/dzmitry-huba/16/base -> origin/gh/dzmitry-huba/16/base
2025-12-04T08:53:09.5348418Z  * [new branch]                gh/dzmitry-huba/16/head -> origin/gh/dzmitry-huba/16/head
2025-12-04T08:53:09.5348496Z  * [new branch]                gh/dzmitry-huba/16/orig -> origin/gh/dzmitry-huba/16/orig
2025-12-04T08:53:09.5348571Z  * [new branch]                gh/dzmitry-huba/17/base -> origin/gh/dzmitry-huba/17/base
2025-12-04T08:53:09.5348645Z  * [new branch]                gh/dzmitry-huba/17/head -> origin/gh/dzmitry-huba/17/head
2025-12-04T08:53:09.5348722Z  * [new branch]                gh/dzmitry-huba/17/orig -> origin/gh/dzmitry-huba/17/orig
2025-12-04T08:53:09.5348798Z  * [new branch]                gh/dzmitry-huba/2/base  -> origin/gh/dzmitry-huba/2/base
2025-12-04T08:53:09.5348875Z  * [new branch]                gh/dzmitry-huba/2/head  -> origin/gh/dzmitry-huba/2/head
2025-12-04T08:53:09.5348953Z  * [new branch]                gh/dzmitry-huba/3/base  -> origin/gh/dzmitry-huba/3/base
2025-12-04T08:53:09.5349028Z  * [new branch]                gh/dzmitry-huba/3/head  -> origin/gh/dzmitry-huba/3/head
2025-12-04T08:53:09.5349103Z  * [new branch]                gh/eellison/808/base    -> origin/gh/eellison/808/base
2025-12-04T08:53:09.5349181Z  * [new branch]                gh/eellison/808/head    -> origin/gh/eellison/808/head
2025-12-04T08:53:09.5349255Z  * [new branch]                gh/eellison/808/orig    -> origin/gh/eellison/808/orig
2025-12-04T08:53:09.5349328Z  * [new branch]                gh/eellison/822/base    -> origin/gh/eellison/822/base
2025-12-04T08:53:09.5349403Z  * [new branch]                gh/eellison/822/head    -> origin/gh/eellison/822/head
2025-12-04T08:53:09.5349504Z  * [new branch]                gh/eellison/822/orig    -> origin/gh/eellison/822/orig
2025-12-04T08:53:09.5349600Z  * [new branch]                gh/eellison/823/base    -> origin/gh/eellison/823/base
2025-12-04T08:53:09.5349671Z  * [new branch]                gh/eellison/823/head    -> origin/gh/eellison/823/head
2025-12-04T08:53:09.5349743Z  * [new branch]                gh/eellison/823/orig    -> origin/gh/eellison/823/orig
2025-12-04T08:53:09.5349817Z  * [new branch]                gh/eellison/862/base    -> origin/gh/eellison/862/base
2025-12-04T08:53:09.5349888Z  * [new branch]                gh/eellison/862/head    -> origin/gh/eellison/862/head
2025-12-04T08:53:09.5349959Z  * [new branch]                gh/eellison/862/orig    -> origin/gh/eellison/862/orig
2025-12-04T08:53:09.5350031Z  * [new branch]                gh/eellison/863/base    -> origin/gh/eellison/863/base
2025-12-04T08:53:09.5350104Z  * [new branch]                gh/eellison/863/head    -> origin/gh/eellison/863/head
2025-12-04T08:53:09.5350176Z  * [new branch]                gh/eellison/863/orig    -> origin/gh/eellison/863/orig
2025-12-04T08:53:09.5350252Z  * [new branch]                gh/eellison/864/base    -> origin/gh/eellison/864/base
2025-12-04T08:53:09.5350323Z  * [new branch]                gh/eellison/864/head    -> origin/gh/eellison/864/head
2025-12-04T08:53:09.5350396Z  * [new branch]                gh/eellison/864/orig    -> origin/gh/eellison/864/orig
2025-12-04T08:53:09.5350499Z  * [new branch]                gh/eellison/865/base    -> origin/gh/eellison/865/base
2025-12-04T08:53:09.5350571Z  * [new branch]                gh/eellison/865/head    -> origin/gh/eellison/865/head
2025-12-04T08:53:09.5350644Z  * [new branch]                gh/eellison/865/orig    -> origin/gh/eellison/865/orig
2025-12-04T08:53:09.5350717Z  * [new branch]                gh/eellison/866/base    -> origin/gh/eellison/866/base
2025-12-04T08:53:09.5350790Z  * [new branch]                gh/eellison/866/head    -> origin/gh/eellison/866/head
2025-12-04T08:53:09.5350866Z  * [new branch]                gh/eellison/866/orig    -> origin/gh/eellison/866/orig
2025-12-04T08:53:09.5350941Z  * [new branch]                gh/eellison/867/base    -> origin/gh/eellison/867/base
2025-12-04T08:53:09.5351012Z  * [new branch]                gh/eellison/867/head    -> origin/gh/eellison/867/head
2025-12-04T08:53:09.5351084Z  * [new branch]                gh/eellison/867/orig    -> origin/gh/eellison/867/orig
2025-12-04T08:53:09.5351161Z  * [new branch]                gh/eellison/868/base    -> origin/gh/eellison/868/base
2025-12-04T08:53:09.5351232Z  * [new branch]                gh/eellison/868/head    -> origin/gh/eellison/868/head
2025-12-04T08:53:09.5351303Z  * [new branch]                gh/eellison/868/orig    -> origin/gh/eellison/868/orig
2025-12-04T08:53:09.5351374Z  * [new branch]                gh/eellison/869/base    -> origin/gh/eellison/869/base
2025-12-04T08:53:09.5351446Z  * [new branch]                gh/eellison/869/head    -> origin/gh/eellison/869/head
2025-12-04T08:53:09.5351519Z  * [new branch]                gh/eellison/869/orig    -> origin/gh/eellison/869/orig
2025-12-04T08:53:09.5351593Z  * [new branch]                gh/eellison/870/base    -> origin/gh/eellison/870/base
2025-12-04T08:53:09.5351665Z  * [new branch]                gh/eellison/870/head    -> origin/gh/eellison/870/head
2025-12-04T08:53:09.5351741Z  * [new branch]                gh/eellison/870/orig    -> origin/gh/eellison/870/orig
2025-12-04T08:53:09.5351812Z  * [new branch]                gh/eellison/871/base    -> origin/gh/eellison/871/base
2025-12-04T08:53:09.5351883Z  * [new branch]                gh/eellison/871/head    -> origin/gh/eellison/871/head
2025-12-04T08:53:09.5351956Z  * [new branch]                gh/eellison/871/orig    -> origin/gh/eellison/871/orig
2025-12-04T08:53:09.5352028Z  * [new branch]                gh/eellison/872/base    -> origin/gh/eellison/872/base
2025-12-04T08:53:09.5352142Z  * [new branch]                gh/eellison/872/head    -> origin/gh/eellison/872/head
2025-12-04T08:53:09.5352215Z  * [new branch]                gh/eellison/872/orig    -> origin/gh/eellison/872/orig
2025-12-04T08:53:09.5352322Z  * [new branch]                gh/eellison/873/base    -> origin/gh/eellison/873/base
2025-12-04T08:53:09.5352393Z  * [new branch]                gh/eellison/873/head    -> origin/gh/eellison/873/head
2025-12-04T08:53:09.5352466Z  * [new branch]                gh/eellison/873/orig    -> origin/gh/eellison/873/orig
2025-12-04T08:53:09.5352537Z  * [new branch]                gh/eellison/874/base    -> origin/gh/eellison/874/base
2025-12-04T08:53:09.5352608Z  * [new branch]                gh/eellison/874/head    -> origin/gh/eellison/874/head
2025-12-04T08:53:09.5352683Z  * [new branch]                gh/eellison/874/orig    -> origin/gh/eellison/874/orig
2025-12-04T08:53:09.5352756Z  * [new branch]                gh/eellison/875/base    -> origin/gh/eellison/875/base
2025-12-04T08:53:09.5352835Z  * [new branch]                gh/eellison/875/head    -> origin/gh/eellison/875/head
2025-12-04T08:53:09.5352911Z  * [new branch]                gh/eellison/875/orig    -> origin/gh/eellison/875/orig
2025-12-04T08:53:09.5352982Z  * [new branch]                gh/eellison/876/base    -> origin/gh/eellison/876/base
2025-12-04T08:53:09.5353055Z  * [new branch]                gh/eellison/876/head    -> origin/gh/eellison/876/head
2025-12-04T08:53:09.5353128Z  * [new branch]                gh/eellison/876/orig    -> origin/gh/eellison/876/orig
2025-12-04T08:53:09.5353201Z  * [new branch]                gh/eellison/877/base    -> origin/gh/eellison/877/base
2025-12-04T08:53:09.5353273Z  * [new branch]                gh/eellison/877/head    -> origin/gh/eellison/877/head
2025-12-04T08:53:09.5353344Z  * [new branch]                gh/eellison/877/orig    -> origin/gh/eellison/877/orig
2025-12-04T08:53:09.5353415Z  * [new branch]                gh/eellison/878/base    -> origin/gh/eellison/878/base
2025-12-04T08:53:09.5353489Z  * [new branch]                gh/eellison/878/head    -> origin/gh/eellison/878/head
2025-12-04T08:53:09.5353561Z  * [new branch]                gh/eellison/878/orig    -> origin/gh/eellison/878/orig
2025-12-04T08:53:09.5353632Z  * [new branch]                gh/eellison/879/base    -> origin/gh/eellison/879/base
2025-12-04T08:53:09.5353704Z  * [new branch]                gh/eellison/879/head    -> origin/gh/eellison/879/head
2025-12-04T08:53:09.5353774Z  * [new branch]                gh/eellison/879/orig    -> origin/gh/eellison/879/orig
2025-12-04T08:53:09.5353847Z  * [new branch]                gh/eellison/880/base    -> origin/gh/eellison/880/base
2025-12-04T08:53:09.5353923Z  * [new branch]                gh/eellison/880/head    -> origin/gh/eellison/880/head
2025-12-04T08:53:09.5353994Z  * [new branch]                gh/eellison/880/orig    -> origin/gh/eellison/880/orig
2025-12-04T08:53:09.5354066Z  * [new branch]                gh/eellison/881/base    -> origin/gh/eellison/881/base
2025-12-04T08:53:09.5354142Z  * [new branch]                gh/eellison/881/head    -> origin/gh/eellison/881/head
2025-12-04T08:53:09.5354216Z  * [new branch]                gh/eellison/881/orig    -> origin/gh/eellison/881/orig
2025-12-04T08:53:09.5354290Z  * [new branch]                gh/eellison/882/base    -> origin/gh/eellison/882/base
2025-12-04T08:53:09.5354361Z  * [new branch]                gh/eellison/882/head    -> origin/gh/eellison/882/head
2025-12-04T08:53:09.5354431Z  * [new branch]                gh/eellison/882/orig    -> origin/gh/eellison/882/orig
2025-12-04T08:53:09.5354507Z  * [new branch]                gh/eellison/883/base    -> origin/gh/eellison/883/base
2025-12-04T08:53:09.5354580Z  * [new branch]                gh/eellison/883/head    -> origin/gh/eellison/883/head
2025-12-04T08:53:09.5354652Z  * [new branch]                gh/eellison/883/orig    -> origin/gh/eellison/883/orig
2025-12-04T08:53:09.5354725Z  * [new branch]                gh/eellison/884/base    -> origin/gh/eellison/884/base
2025-12-04T08:53:09.5354820Z  * [new branch]                gh/eellison/884/head    -> origin/gh/eellison/884/head
2025-12-04T08:53:09.5355287Z  * [new branch]                gh/eellison/884/orig    -> origin/gh/eellison/884/orig
2025-12-04T08:53:09.5355359Z  * [new branch]                gh/etaf/147/base        -> origin/gh/etaf/147/base
2025-12-04T08:53:09.5355426Z  * [new branch]                gh/etaf/147/head        -> origin/gh/etaf/147/head
2025-12-04T08:53:09.5355490Z  * [new branch]                gh/etaf/154/base        -> origin/gh/etaf/154/base
2025-12-04T08:53:09.5355559Z  * [new branch]                gh/etaf/154/head        -> origin/gh/etaf/154/head
2025-12-04T08:53:09.5355626Z  * [new branch]                gh/etaf/154/orig        -> origin/gh/etaf/154/orig
2025-12-04T08:53:09.5355691Z  * [new branch]                gh/etaf/156/base        -> origin/gh/etaf/156/base
2025-12-04T08:53:09.5355758Z  * [new branch]                gh/etaf/156/head        -> origin/gh/etaf/156/head
2025-12-04T08:53:09.5355824Z  * [new branch]                gh/etaf/156/orig        -> origin/gh/etaf/156/orig
2025-12-04T08:53:09.5355889Z  * [new branch]                gh/etaf/157/base        -> origin/gh/etaf/157/base
2025-12-04T08:53:09.5355956Z  * [new branch]                gh/etaf/157/head        -> origin/gh/etaf/157/head
2025-12-04T08:53:09.5356020Z  * [new branch]                gh/etaf/157/orig        -> origin/gh/etaf/157/orig
2025-12-04T08:53:09.5356084Z  * [new branch]                gh/etaf/158/base        -> origin/gh/etaf/158/base
2025-12-04T08:53:09.5356151Z  * [new branch]                gh/etaf/158/head        -> origin/gh/etaf/158/head
2025-12-04T08:53:09.5356216Z  * [new branch]                gh/etaf/158/orig        -> origin/gh/etaf/158/orig
2025-12-04T08:53:09.5356286Z  * [new branch]                gh/etaf/159/base        -> origin/gh/etaf/159/base
2025-12-04T08:53:09.5356351Z  * [new branch]                gh/etaf/159/head        -> origin/gh/etaf/159/head
2025-12-04T08:53:09.5356418Z  * [new branch]                gh/etaf/159/orig        -> origin/gh/etaf/159/orig
2025-12-04T08:53:09.5356486Z  * [new branch]                gh/etaf/160/base        -> origin/gh/etaf/160/base
2025-12-04T08:53:09.5356556Z  * [new branch]                gh/etaf/160/head        -> origin/gh/etaf/160/head
2025-12-04T08:53:09.5356620Z  * [new branch]                gh/etaf/160/orig        -> origin/gh/etaf/160/orig
2025-12-04T08:53:09.5356688Z  * [new branch]                gh/etaf/161/base        -> origin/gh/etaf/161/base
2025-12-04T08:53:09.5356753Z  * [new branch]                gh/etaf/161/head        -> origin/gh/etaf/161/head
2025-12-04T08:53:09.5356819Z  * [new branch]                gh/etaf/161/orig        -> origin/gh/etaf/161/orig
2025-12-04T08:53:09.5356887Z  * [new branch]                gh/etaf/166/base        -> origin/gh/etaf/166/base
2025-12-04T08:53:09.5356952Z  * [new branch]                gh/etaf/166/head        -> origin/gh/etaf/166/head
2025-12-04T08:53:09.5357020Z  * [new branch]                gh/etaf/166/orig        -> origin/gh/etaf/166/orig
2025-12-04T08:53:09.5357088Z  * [new branch]                gh/etaf/167/base        -> origin/gh/etaf/167/base
2025-12-04T08:53:09.5357153Z  * [new branch]                gh/etaf/167/head        -> origin/gh/etaf/167/head
2025-12-04T08:53:09.5357220Z  * [new branch]                gh/etaf/167/orig        -> origin/gh/etaf/167/orig
2025-12-04T08:53:09.5357287Z  * [new branch]                gh/etaf/168/base        -> origin/gh/etaf/168/base
2025-12-04T08:53:09.5357351Z  * [new branch]                gh/etaf/168/head        -> origin/gh/etaf/168/head
2025-12-04T08:53:09.5357418Z  * [new branch]                gh/etaf/168/orig        -> origin/gh/etaf/168/orig
2025-12-04T08:53:09.5357486Z  * [new branch]                gh/etaf/172/base        -> origin/gh/etaf/172/base
2025-12-04T08:53:09.5357550Z  * [new branch]                gh/etaf/172/head        -> origin/gh/etaf/172/head
2025-12-04T08:53:09.5357616Z  * [new branch]                gh/etaf/172/orig        -> origin/gh/etaf/172/orig
2025-12-04T08:53:09.5357714Z  * [new branch]                gh/etaf/173/base        -> origin/gh/etaf/173/base
2025-12-04T08:53:09.5357802Z  * [new branch]                gh/etaf/173/head        -> origin/gh/etaf/173/head
2025-12-04T08:53:09.5357869Z  * [new branch]                gh/etaf/173/orig        -> origin/gh/etaf/173/orig
2025-12-04T08:53:09.5357934Z  * [new branch]                gh/etaf/174/base        -> origin/gh/etaf/174/base
2025-12-04T08:53:09.5357999Z  * [new branch]                gh/etaf/174/head        -> origin/gh/etaf/174/head
2025-12-04T08:53:09.5358065Z  * [new branch]                gh/etaf/175/base        -> origin/gh/etaf/175/base
2025-12-04T08:53:09.5358130Z  * [new branch]                gh/etaf/175/head        -> origin/gh/etaf/175/head
2025-12-04T08:53:09.5358194Z  * [new branch]                gh/etaf/175/orig        -> origin/gh/etaf/175/orig
2025-12-04T08:53:09.5358260Z  * [new branch]                gh/etaf/176/base        -> origin/gh/etaf/176/base
2025-12-04T08:53:09.5358326Z  * [new branch]                gh/etaf/176/head        -> origin/gh/etaf/176/head
2025-12-04T08:53:09.5358391Z  * [new branch]                gh/etaf/176/orig        -> origin/gh/etaf/176/orig
2025-12-04T08:53:09.5358458Z  * [new branch]                gh/etaf/177/base        -> origin/gh/etaf/177/base
2025-12-04T08:53:09.5358524Z  * [new branch]                gh/etaf/177/head        -> origin/gh/etaf/177/head
2025-12-04T08:53:09.5358590Z  * [new branch]                gh/etaf/177/orig        -> origin/gh/etaf/177/orig
2025-12-04T08:53:09.5358656Z  * [new branch]                gh/etaf/178/base        -> origin/gh/etaf/178/base
2025-12-04T08:53:09.5358721Z  * [new branch]                gh/etaf/178/head        -> origin/gh/etaf/178/head
2025-12-04T08:53:09.5358785Z  * [new branch]                gh/etaf/178/orig        -> origin/gh/etaf/178/orig
2025-12-04T08:53:09.5358853Z  * [new branch]                gh/etaf/179/base        -> origin/gh/etaf/179/base
2025-12-04T08:53:09.5358921Z  * [new branch]                gh/etaf/179/head        -> origin/gh/etaf/179/head
2025-12-04T08:53:09.5358988Z  * [new branch]                gh/etaf/179/orig        -> origin/gh/etaf/179/orig
2025-12-04T08:53:09.5359055Z  * [new branch]                gh/etaf/180/base        -> origin/gh/etaf/180/base
2025-12-04T08:53:09.5359121Z  * [new branch]                gh/etaf/180/head        -> origin/gh/etaf/180/head
2025-12-04T08:53:09.5359187Z  * [new branch]                gh/etaf/180/orig        -> origin/gh/etaf/180/orig
2025-12-04T08:53:09.5359268Z  * [new branch]                gh/exclamaforte/1/base  -> origin/gh/exclamaforte/1/base
2025-12-04T08:53:09.5359346Z  * [new branch]                gh/exclamaforte/1/head  -> origin/gh/exclamaforte/1/head
2025-12-04T08:53:09.5359422Z  * [new branch]                gh/exclamaforte/2/base  -> origin/gh/exclamaforte/2/base
2025-12-04T08:53:09.5359501Z  * [new branch]                gh/exclamaforte/2/head  -> origin/gh/exclamaforte/2/head
2025-12-04T08:53:09.5359578Z  * [new branch]                gh/exclamaforte/3/base  -> origin/gh/exclamaforte/3/base
2025-12-04T08:53:09.5359659Z  * [new branch]                gh/exclamaforte/3/head  -> origin/gh/exclamaforte/3/head
2025-12-04T08:53:09.5359734Z  * [new branch]                gh/exclamaforte/4/base  -> origin/gh/exclamaforte/4/base
2025-12-04T08:53:09.5359810Z  * [new branch]                gh/exclamaforte/4/head  -> origin/gh/exclamaforte/4/head
2025-12-04T08:53:09.5359883Z  * [new branch]                gh/ezyang/2374/base     -> origin/gh/ezyang/2374/base
2025-12-04T08:53:09.5359954Z  * [new branch]                gh/ezyang/2374/head     -> origin/gh/ezyang/2374/head
2025-12-04T08:53:09.5360024Z  * [new branch]                gh/ezyang/2374/orig     -> origin/gh/ezyang/2374/orig
2025-12-04T08:53:09.5360096Z  * [new branch]                gh/ezyang/2973/base     -> origin/gh/ezyang/2973/base
2025-12-04T08:53:09.5360165Z  * [new branch]                gh/ezyang/2973/head     -> origin/gh/ezyang/2973/head
2025-12-04T08:53:09.5360276Z  * [new branch]                gh/ezyang/2973/orig     -> origin/gh/ezyang/2973/orig
2025-12-04T08:53:09.5360378Z  * [new branch]                gh/ezyang/2974/base     -> origin/gh/ezyang/2974/base
2025-12-04T08:53:09.5360477Z  * [new branch]                gh/ezyang/2974/head     -> origin/gh/ezyang/2974/head
2025-12-04T08:53:09.5360546Z  * [new branch]                gh/ezyang/2974/orig     -> origin/gh/ezyang/2974/orig
2025-12-04T08:53:09.5360616Z  * [new branch]                gh/ezyang/3131/base     -> origin/gh/ezyang/3131/base
2025-12-04T08:53:09.5360684Z  * [new branch]                gh/ezyang/3131/head     -> origin/gh/ezyang/3131/head
2025-12-04T08:53:09.5360751Z  * [new branch]                gh/ezyang/3131/orig     -> origin/gh/ezyang/3131/orig
2025-12-04T08:53:09.5360822Z  * [new branch]                gh/ezyang/3139/base     -> origin/gh/ezyang/3139/base
2025-12-04T08:53:09.5360890Z  * [new branch]                gh/ezyang/3139/head     -> origin/gh/ezyang/3139/head
2025-12-04T08:53:09.5360961Z  * [new branch]                gh/ezyang/3139/orig     -> origin/gh/ezyang/3139/orig
2025-12-04T08:53:09.5361039Z  * [new branch]                gh/ezyang/3140/base     -> origin/gh/ezyang/3140/base
2025-12-04T08:53:09.5361108Z  * [new branch]                gh/ezyang/3140/head     -> origin/gh/ezyang/3140/head
2025-12-04T08:53:09.5361176Z  * [new branch]                gh/ezyang/3140/orig     -> origin/gh/ezyang/3140/orig
2025-12-04T08:53:09.5361249Z  * [new branch]                gh/ezyang/3143/base     -> origin/gh/ezyang/3143/base
2025-12-04T08:53:09.5361317Z  * [new branch]                gh/ezyang/3143/head     -> origin/gh/ezyang/3143/head
2025-12-04T08:53:09.5361386Z  * [new branch]                gh/ezyang/3143/orig     -> origin/gh/ezyang/3143/orig
2025-12-04T08:53:09.5361455Z  * [new branch]                gh/ezyang/3144/base     -> origin/gh/ezyang/3144/base
2025-12-04T08:53:09.5361523Z  * [new branch]                gh/ezyang/3144/head     -> origin/gh/ezyang/3144/head
2025-12-04T08:53:09.5361598Z  * [new branch]                gh/ezyang/3144/orig     -> origin/gh/ezyang/3144/orig
2025-12-04T08:53:09.5361671Z  * [new branch]                gh/ezyang/3167/base     -> origin/gh/ezyang/3167/base
2025-12-04T08:53:09.5361740Z  * [new branch]                gh/ezyang/3167/head     -> origin/gh/ezyang/3167/head
2025-12-04T08:53:09.5361810Z  * [new branch]                gh/ezyang/3167/orig     -> origin/gh/ezyang/3167/orig
2025-12-04T08:53:09.5361878Z  * [new branch]                gh/ezyang/3173/base     -> origin/gh/ezyang/3173/base
2025-12-04T08:53:09.5361947Z  * [new branch]                gh/ezyang/3173/head     -> origin/gh/ezyang/3173/head
2025-12-04T08:53:09.5362018Z  * [new branch]                gh/ezyang/3173/orig     -> origin/gh/ezyang/3173/orig
2025-12-04T08:53:09.5362087Z  * [new branch]                gh/ezyang/3175/base     -> origin/gh/ezyang/3175/base
2025-12-04T08:53:09.5362157Z  * [new branch]                gh/ezyang/3175/head     -> origin/gh/ezyang/3175/head
2025-12-04T08:53:09.5362232Z  * [new branch]                gh/ezyang/3175/orig     -> origin/gh/ezyang/3175/orig
2025-12-04T08:53:09.5362301Z  * [new branch]                gh/ezyang/3182/base     -> origin/gh/ezyang/3182/base
2025-12-04T08:53:09.5362370Z  * [new branch]                gh/ezyang/3182/head     -> origin/gh/ezyang/3182/head
2025-12-04T08:53:09.5362442Z  * [new branch]                gh/ezyang/3182/orig     -> origin/gh/ezyang/3182/orig
2025-12-04T08:53:09.5362510Z  * [new branch]                gh/ezyang/3185/base     -> origin/gh/ezyang/3185/base
2025-12-04T08:53:09.5362578Z  * [new branch]                gh/ezyang/3185/head     -> origin/gh/ezyang/3185/head
2025-12-04T08:53:09.5362648Z  * [new branch]                gh/ezyang/3185/orig     -> origin/gh/ezyang/3185/orig
2025-12-04T08:53:09.5362717Z  * [new branch]                gh/ezyang/3189/base     -> origin/gh/ezyang/3189/base
2025-12-04T08:53:09.5362785Z  * [new branch]                gh/ezyang/3189/head     -> origin/gh/ezyang/3189/head
2025-12-04T08:53:09.5362901Z  * [new branch]                gh/ezyang/3189/orig     -> origin/gh/ezyang/3189/orig
2025-12-04T08:53:09.5363014Z  * [new branch]                gh/ezyang/3191/base     -> origin/gh/ezyang/3191/base
2025-12-04T08:53:09.5363086Z  * [new branch]                gh/ezyang/3191/head     -> origin/gh/ezyang/3191/head
2025-12-04T08:53:09.5363155Z  * [new branch]                gh/ezyang/3191/orig     -> origin/gh/ezyang/3191/orig
2025-12-04T08:53:09.5363224Z  * [new branch]                gh/ezyang/3192/base     -> origin/gh/ezyang/3192/base
2025-12-04T08:53:09.5363296Z  * [new branch]                gh/ezyang/3192/head     -> origin/gh/ezyang/3192/head
2025-12-04T08:53:09.5363364Z  * [new branch]                gh/ezyang/3192/orig     -> origin/gh/ezyang/3192/orig
2025-12-04T08:53:09.5363432Z  * [new branch]                gh/ezyang/3193/base     -> origin/gh/ezyang/3193/base
2025-12-04T08:53:09.5363501Z  * [new branch]                gh/ezyang/3193/head     -> origin/gh/ezyang/3193/head
2025-12-04T08:53:09.5363572Z  * [new branch]                gh/ezyang/3193/orig     -> origin/gh/ezyang/3193/orig
2025-12-04T08:53:09.5363642Z  * [new branch]                gh/ezyang/3194/base     -> origin/gh/ezyang/3194/base
2025-12-04T08:53:09.5363713Z  * [new branch]                gh/ezyang/3194/head     -> origin/gh/ezyang/3194/head
2025-12-04T08:53:09.5363783Z  * [new branch]                gh/ezyang/3194/orig     -> origin/gh/ezyang/3194/orig
2025-12-04T08:53:09.5363853Z  * [new branch]                gh/ezyang/3195/base     -> origin/gh/ezyang/3195/base
2025-12-04T08:53:09.5363924Z  * [new branch]                gh/ezyang/3195/head     -> origin/gh/ezyang/3195/head
2025-12-04T08:53:09.5363992Z  * [new branch]                gh/ezyang/3195/orig     -> origin/gh/ezyang/3195/orig
2025-12-04T08:53:09.5364308Z  * [new branch]                gh/ezyang/3196/base     -> origin/gh/ezyang/3196/base
2025-12-04T08:53:09.5364382Z  * [new branch]                gh/ezyang/3196/head     -> origin/gh/ezyang/3196/head
2025-12-04T08:53:09.5364455Z  * [new branch]                gh/ezyang/3196/orig     -> origin/gh/ezyang/3196/orig
2025-12-04T08:53:09.5364530Z  * [new branch]                gh/ezyang/3197/base     -> origin/gh/ezyang/3197/base
2025-12-04T08:53:09.5364601Z  * [new branch]                gh/ezyang/3197/head     -> origin/gh/ezyang/3197/head
2025-12-04T08:53:09.5364670Z  * [new branch]                gh/ezyang/3197/orig     -> origin/gh/ezyang/3197/orig
2025-12-04T08:53:09.5364740Z  * [new branch]                gh/ezyang/3198/base     -> origin/gh/ezyang/3198/base
2025-12-04T08:53:09.5364810Z  * [new branch]                gh/ezyang/3198/head     -> origin/gh/ezyang/3198/head
2025-12-04T08:53:09.5364880Z  * [new branch]                gh/ezyang/3198/orig     -> origin/gh/ezyang/3198/orig
2025-12-04T08:53:09.5364952Z  * [new branch]                gh/ezyang/3199/base     -> origin/gh/ezyang/3199/base
2025-12-04T08:53:09.5365022Z  * [new branch]                gh/ezyang/3199/head     -> origin/gh/ezyang/3199/head
2025-12-04T08:53:09.5365094Z  * [new branch]                gh/ezyang/3199/orig     -> origin/gh/ezyang/3199/orig
2025-12-04T08:53:09.5365165Z  * [new branch]                gh/ezyang/3200/base     -> origin/gh/ezyang/3200/base
2025-12-04T08:53:09.5365233Z  * [new branch]                gh/ezyang/3200/head     -> origin/gh/ezyang/3200/head
2025-12-04T08:53:09.5365301Z  * [new branch]                gh/ezyang/3200/orig     -> origin/gh/ezyang/3200/orig
2025-12-04T08:53:09.5365370Z  * [new branch]                gh/ezyang/3201/base     -> origin/gh/ezyang/3201/base
2025-12-04T08:53:09.5365439Z  * [new branch]                gh/ezyang/3201/head     -> origin/gh/ezyang/3201/head
2025-12-04T08:53:09.5365509Z  * [new branch]                gh/ezyang/3201/orig     -> origin/gh/ezyang/3201/orig
2025-12-04T08:53:09.5365580Z  * [new branch]                gh/ezyang/3202/base     -> origin/gh/ezyang/3202/base
2025-12-04T08:53:09.5365649Z  * [new branch]                gh/ezyang/3202/head     -> origin/gh/ezyang/3202/head
2025-12-04T08:53:09.5365745Z  * [new branch]                gh/ezyang/3202/orig     -> origin/gh/ezyang/3202/orig
2025-12-04T08:53:09.5365843Z  * [new branch]                gh/ezyang/3203/base     -> origin/gh/ezyang/3203/base
2025-12-04T08:53:09.5365911Z  * [new branch]                gh/ezyang/3203/head     -> origin/gh/ezyang/3203/head
2025-12-04T08:53:09.5365980Z  * [new branch]                gh/ezyang/3203/orig     -> origin/gh/ezyang/3203/orig
2025-12-04T08:53:09.5366050Z  * [new branch]                gh/ezyang/3204/base     -> origin/gh/ezyang/3204/base
2025-12-04T08:53:09.5366120Z  * [new branch]                gh/ezyang/3204/head     -> origin/gh/ezyang/3204/head
2025-12-04T08:53:09.5366190Z  * [new branch]                gh/ezyang/3204/orig     -> origin/gh/ezyang/3204/orig
2025-12-04T08:53:09.5366267Z  * [new branch]                gh/ezyang/3205/base     -> origin/gh/ezyang/3205/base
2025-12-04T08:53:09.5372665Z  * [new branch]                gh/ezyang/3205/head     -> origin/gh/ezyang/3205/head
2025-12-04T08:53:09.5372761Z  * [new branch]                gh/ezyang/3205/orig     -> origin/gh/ezyang/3205/orig
2025-12-04T08:53:09.5372842Z  * [new branch]                gh/ezyang/3206/base     -> origin/gh/ezyang/3206/base
2025-12-04T08:53:09.5372915Z  * [new branch]                gh/ezyang/3206/head     -> origin/gh/ezyang/3206/head
2025-12-04T08:53:09.5372988Z  * [new branch]                gh/ezyang/3206/orig     -> origin/gh/ezyang/3206/orig
2025-12-04T08:53:09.5373061Z  * [new branch]                gh/ezyang/3207/base     -> origin/gh/ezyang/3207/base
2025-12-04T08:53:09.5373129Z  * [new branch]                gh/ezyang/3207/head     -> origin/gh/ezyang/3207/head
2025-12-04T08:53:09.5373200Z  * [new branch]                gh/ezyang/3207/orig     -> origin/gh/ezyang/3207/orig
2025-12-04T08:53:09.5373272Z  * [new branch]                gh/ezyang/3208/base     -> origin/gh/ezyang/3208/base
2025-12-04T08:53:09.5373339Z  * [new branch]                gh/ezyang/3208/head     -> origin/gh/ezyang/3208/head
2025-12-04T08:53:09.5373410Z  * [new branch]                gh/ezyang/3208/orig     -> origin/gh/ezyang/3208/orig
2025-12-04T08:53:09.5373481Z  * [new branch]                gh/ezyang/3209/base     -> origin/gh/ezyang/3209/base
2025-12-04T08:53:09.5373553Z  * [new branch]                gh/ezyang/3209/head     -> origin/gh/ezyang/3209/head
2025-12-04T08:53:09.5373624Z  * [new branch]                gh/ezyang/3209/orig     -> origin/gh/ezyang/3209/orig
2025-12-04T08:53:09.5373699Z  * [new branch]                gh/fadara01/3/base      -> origin/gh/fadara01/3/base
2025-12-04T08:53:09.5373771Z  * [new branch]                gh/fadara01/3/head      -> origin/gh/fadara01/3/head
2025-12-04T08:53:09.5373848Z  * [new branch]                gh/fadara01/3/orig      -> origin/gh/fadara01/3/orig
2025-12-04T08:53:09.5373919Z  * [new branch]                gh/fadara01/5/base      -> origin/gh/fadara01/5/base
2025-12-04T08:53:09.5373990Z  * [new branch]                gh/fadara01/5/head      -> origin/gh/fadara01/5/head
2025-12-04T08:53:09.5374063Z  * [new branch]                gh/fadara01/5/orig      -> origin/gh/fadara01/5/orig
2025-12-04T08:53:09.5374134Z  * [new branch]                gh/fadara01/6/base      -> origin/gh/fadara01/6/base
2025-12-04T08:53:09.5374206Z  * [new branch]                gh/fadara01/6/head      -> origin/gh/fadara01/6/head
2025-12-04T08:53:09.5374278Z  * [new branch]                gh/fadara01/6/orig      -> origin/gh/fadara01/6/orig
2025-12-04T08:53:09.5374347Z  * [new branch]                gh/fadara01/7/base      -> origin/gh/fadara01/7/base
2025-12-04T08:53:09.5374415Z  * [new branch]                gh/fadara01/7/head      -> origin/gh/fadara01/7/head
2025-12-04T08:53:09.5374484Z  * [new branch]                gh/fadara01/7/orig      -> origin/gh/fadara01/7/orig
2025-12-04T08:53:09.5374553Z  * [new branch]                gh/fadara01/8/base      -> origin/gh/fadara01/8/base
2025-12-04T08:53:09.5374625Z  * [new branch]                gh/fadara01/8/head      -> origin/gh/fadara01/8/head
2025-12-04T08:53:09.5374766Z  * [new branch]                gh/fadara01/8/orig      -> origin/gh/fadara01/8/orig
2025-12-04T08:53:09.5374887Z  * [new branch]                gh/fadara01/9/base      -> origin/gh/fadara01/9/base
2025-12-04T08:53:09.5374963Z  * [new branch]                gh/fadara01/9/head      -> origin/gh/fadara01/9/head
2025-12-04T08:53:09.5375031Z  * [new branch]                gh/fadara01/9/orig      -> origin/gh/fadara01/9/orig
2025-12-04T08:53:09.5375101Z  * [new branch]                gh/fduwjj/182/base      -> origin/gh/fduwjj/182/base
2025-12-04T08:53:09.5375172Z  * [new branch]                gh/fduwjj/182/head      -> origin/gh/fduwjj/182/head
2025-12-04T08:53:09.5375245Z  * [new branch]                gh/fduwjj/182/orig      -> origin/gh/fduwjj/182/orig
2025-12-04T08:53:09.5375314Z  * [new branch]                gh/fduwjj/211/base      -> origin/gh/fduwjj/211/base
2025-12-04T08:53:09.5375382Z  * [new branch]                gh/fduwjj/211/head      -> origin/gh/fduwjj/211/head
2025-12-04T08:53:09.5375454Z  * [new branch]                gh/fduwjj/211/orig      -> origin/gh/fduwjj/211/orig
2025-12-04T08:53:09.5375529Z  * [new branch]                gh/fduwjj/212/base      -> origin/gh/fduwjj/212/base
2025-12-04T08:53:09.5375600Z  * [new branch]                gh/fduwjj/212/head      -> origin/gh/fduwjj/212/head
2025-12-04T08:53:09.5375668Z  * [new branch]                gh/fduwjj/212/orig      -> origin/gh/fduwjj/212/orig
2025-12-04T08:53:09.5375738Z  * [new branch]                gh/fduwjj/213/base      -> origin/gh/fduwjj/213/base
2025-12-04T08:53:09.5375814Z  * [new branch]                gh/fduwjj/213/head      -> origin/gh/fduwjj/213/head
2025-12-04T08:53:09.5375883Z  * [new branch]                gh/fduwjj/213/orig      -> origin/gh/fduwjj/213/orig
2025-12-04T08:53:09.5375950Z  * [new branch]                gh/fduwjj/226/base      -> origin/gh/fduwjj/226/base
2025-12-04T08:53:09.5376024Z  * [new branch]                gh/fduwjj/226/head      -> origin/gh/fduwjj/226/head
2025-12-04T08:53:09.5376097Z  * [new branch]                gh/fduwjj/226/orig      -> origin/gh/fduwjj/226/orig
2025-12-04T08:53:09.5376167Z  * [new branch]                gh/fduwjj/229/base      -> origin/gh/fduwjj/229/base
2025-12-04T08:53:09.5376237Z  * [new branch]                gh/fduwjj/229/head      -> origin/gh/fduwjj/229/head
2025-12-04T08:53:09.5376306Z  * [new branch]                gh/fduwjj/229/orig      -> origin/gh/fduwjj/229/orig
2025-12-04T08:53:09.5376379Z  * [new branch]                gh/fduwjj/233/base      -> origin/gh/fduwjj/233/base
2025-12-04T08:53:09.5376452Z  * [new branch]                gh/fduwjj/233/head      -> origin/gh/fduwjj/233/head
2025-12-04T08:53:09.5376521Z  * [new branch]                gh/fduwjj/233/orig      -> origin/gh/fduwjj/233/orig
2025-12-04T08:53:09.5376593Z  * [new branch]                gh/fduwjj/234/base      -> origin/gh/fduwjj/234/base
2025-12-04T08:53:09.5376665Z  * [new branch]                gh/fduwjj/234/head      -> origin/gh/fduwjj/234/head
2025-12-04T08:53:09.5376736Z  * [new branch]                gh/fduwjj/234/orig      -> origin/gh/fduwjj/234/orig
2025-12-04T08:53:09.5376814Z  * [new branch]                gh/fduwjj/235/base      -> origin/gh/fduwjj/235/base
2025-12-04T08:53:09.5376883Z  * [new branch]                gh/fduwjj/235/head      -> origin/gh/fduwjj/235/head
2025-12-04T08:53:09.5376951Z  * [new branch]                gh/fduwjj/235/orig      -> origin/gh/fduwjj/235/orig
2025-12-04T08:53:09.5377021Z  * [new branch]                gh/fduwjj/236/base      -> origin/gh/fduwjj/236/base
2025-12-04T08:53:09.5377094Z  * [new branch]                gh/fduwjj/236/head      -> origin/gh/fduwjj/236/head
2025-12-04T08:53:09.5377163Z  * [new branch]                gh/fduwjj/236/orig      -> origin/gh/fduwjj/236/orig
2025-12-04T08:53:09.5377234Z  * [new branch]                gh/fduwjj/237/base      -> origin/gh/fduwjj/237/base
2025-12-04T08:53:09.5377301Z  * [new branch]                gh/fduwjj/237/head      -> origin/gh/fduwjj/237/head
2025-12-04T08:53:09.5377393Z  * [new branch]                gh/fduwjj/237/orig      -> origin/gh/fduwjj/237/orig
2025-12-04T08:53:09.5377467Z  * [new branch]                gh/fduwjj/238/base      -> origin/gh/fduwjj/238/base
2025-12-04T08:53:09.5377565Z  * [new branch]                gh/fduwjj/238/head      -> origin/gh/fduwjj/238/head
2025-12-04T08:53:09.5377635Z  * [new branch]                gh/fduwjj/238/orig      -> origin/gh/fduwjj/238/orig
2025-12-04T08:53:09.5377710Z  * [new branch]                gh/fduwjj/239/base      -> origin/gh/fduwjj/239/base
2025-12-04T08:53:09.5377777Z  * [new branch]                gh/fduwjj/239/head      -> origin/gh/fduwjj/239/head
2025-12-04T08:53:09.5377845Z  * [new branch]                gh/fduwjj/239/orig      -> origin/gh/fduwjj/239/orig
2025-12-04T08:53:09.5377918Z  * [new branch]                gh/fegin/332/base       -> origin/gh/fegin/332/base
2025-12-04T08:53:09.5377990Z  * [new branch]                gh/fegin/332/head       -> origin/gh/fegin/332/head
2025-12-04T08:53:09.5378059Z  * [new branch]                gh/fegin/332/orig       -> origin/gh/fegin/332/orig
2025-12-04T08:53:09.5378125Z  * [new branch]                gh/fegin/333/base       -> origin/gh/fegin/333/base
2025-12-04T08:53:09.5378195Z  * [new branch]                gh/fegin/333/head       -> origin/gh/fegin/333/head
2025-12-04T08:53:09.5378260Z  * [new branch]                gh/fegin/333/orig       -> origin/gh/fegin/333/orig
2025-12-04T08:53:09.5378325Z  * [new branch]                gh/fegin/334/base       -> origin/gh/fegin/334/base
2025-12-04T08:53:09.5378391Z  * [new branch]                gh/fegin/334/head       -> origin/gh/fegin/334/head
2025-12-04T08:53:09.5378457Z  * [new branch]                gh/fegin/334/orig       -> origin/gh/fegin/334/orig
2025-12-04T08:53:09.5378522Z  * [new branch]                gh/fegin/335/base       -> origin/gh/fegin/335/base
2025-12-04T08:53:09.5378588Z  * [new branch]                gh/fegin/335/head       -> origin/gh/fegin/335/head
2025-12-04T08:53:09.5378655Z  * [new branch]                gh/fegin/335/orig       -> origin/gh/fegin/335/orig
2025-12-04T08:53:09.5378723Z  * [new branch]                gh/fffrog/160/base      -> origin/gh/fffrog/160/base
2025-12-04T08:53:09.5378795Z  * [new branch]                gh/fffrog/160/head      -> origin/gh/fffrog/160/head
2025-12-04T08:53:09.5378863Z  * [new branch]                gh/fffrog/177/base      -> origin/gh/fffrog/177/base
2025-12-04T08:53:09.5378931Z  * [new branch]                gh/fffrog/177/head      -> origin/gh/fffrog/177/head
2025-12-04T08:53:09.5378998Z  * [new branch]                gh/fffrog/177/orig      -> origin/gh/fffrog/177/orig
2025-12-04T08:53:09.5379064Z  * [new branch]                gh/fffrog/178/base      -> origin/gh/fffrog/178/base
2025-12-04T08:53:09.5379132Z  * [new branch]                gh/fffrog/178/head      -> origin/gh/fffrog/178/head
2025-12-04T08:53:09.5379199Z  * [new branch]                gh/fffrog/178/orig      -> origin/gh/fffrog/178/orig
2025-12-04T08:53:09.5379267Z  * [new branch]                gh/fffrog/181/base      -> origin/gh/fffrog/181/base
2025-12-04T08:53:09.5379335Z  * [new branch]                gh/fffrog/181/head      -> origin/gh/fffrog/181/head
2025-12-04T08:53:09.5379404Z  * [new branch]                gh/fffrog/181/orig      -> origin/gh/fffrog/181/orig
2025-12-04T08:53:09.5379470Z  * [new branch]                gh/fffrog/183/base      -> origin/gh/fffrog/183/base
2025-12-04T08:53:09.5379538Z  * [new branch]                gh/fffrog/183/head      -> origin/gh/fffrog/183/head
2025-12-04T08:53:09.5379604Z  * [new branch]                gh/fffrog/183/orig      -> origin/gh/fffrog/183/orig
2025-12-04T08:53:09.5379673Z  * [new branch]                gh/fxdawnn/10/base      -> origin/gh/fxdawnn/10/base
2025-12-04T08:53:09.5379741Z  * [new branch]                gh/fxdawnn/10/head      -> origin/gh/fxdawnn/10/head
2025-12-04T08:53:09.5379809Z  * [new branch]                gh/fxdawnn/10/orig      -> origin/gh/fxdawnn/10/orig
2025-12-04T08:53:09.5379903Z  * [new branch]                gh/fxdawnn/11/base      -> origin/gh/fxdawnn/11/base
2025-12-04T08:53:09.5379971Z  * [new branch]                gh/fxdawnn/11/head      -> origin/gh/fxdawnn/11/head
2025-12-04T08:53:09.5380063Z  * [new branch]                gh/fxdawnn/11/orig      -> origin/gh/fxdawnn/11/orig
2025-12-04T08:53:09.5380131Z  * [new branch]                gh/fxdawnn/12/base      -> origin/gh/fxdawnn/12/base
2025-12-04T08:53:09.5380199Z  * [new branch]                gh/fxdawnn/12/head      -> origin/gh/fxdawnn/12/head
2025-12-04T08:53:09.5380265Z  * [new branch]                gh/fxdawnn/12/orig      -> origin/gh/fxdawnn/12/orig
2025-12-04T08:53:09.5380332Z  * [new branch]                gh/fxdawnn/13/base      -> origin/gh/fxdawnn/13/base
2025-12-04T08:53:09.5380401Z  * [new branch]                gh/fxdawnn/13/head      -> origin/gh/fxdawnn/13/head
2025-12-04T08:53:09.5380498Z  * [new branch]                gh/fxdawnn/13/orig      -> origin/gh/fxdawnn/13/orig
2025-12-04T08:53:09.5380567Z  * [new branch]                gh/fxdawnn/14/base      -> origin/gh/fxdawnn/14/base
2025-12-04T08:53:09.5380636Z  * [new branch]                gh/fxdawnn/14/head      -> origin/gh/fxdawnn/14/head
2025-12-04T08:53:09.5380704Z  * [new branch]                gh/fxdawnn/14/orig      -> origin/gh/fxdawnn/14/orig
2025-12-04T08:53:09.5380772Z  * [new branch]                gh/fxdawnn/15/base      -> origin/gh/fxdawnn/15/base
2025-12-04T08:53:09.5380840Z  * [new branch]                gh/fxdawnn/15/head      -> origin/gh/fxdawnn/15/head
2025-12-04T08:53:09.5380907Z  * [new branch]                gh/fxdawnn/15/orig      -> origin/gh/fxdawnn/15/orig
2025-12-04T08:53:09.5380976Z  * [new branch]                gh/fxdawnn/6/base       -> origin/gh/fxdawnn/6/base
2025-12-04T08:53:09.5381042Z  * [new branch]                gh/fxdawnn/6/head       -> origin/gh/fxdawnn/6/head
2025-12-04T08:53:09.5381109Z  * [new branch]                gh/fxdawnn/6/orig       -> origin/gh/fxdawnn/6/orig
2025-12-04T08:53:09.5381176Z  * [new branch]                gh/fxdawnn/7/base       -> origin/gh/fxdawnn/7/base
2025-12-04T08:53:09.5381244Z  * [new branch]                gh/fxdawnn/7/head       -> origin/gh/fxdawnn/7/head
2025-12-04T08:53:09.5381311Z  * [new branch]                gh/fxdawnn/7/orig       -> origin/gh/fxdawnn/7/orig
2025-12-04T08:53:09.5381381Z  * [new branch]                gh/fxdawnn/9/base       -> origin/gh/fxdawnn/9/base
2025-12-04T08:53:09.5381449Z  * [new branch]                gh/fxdawnn/9/head       -> origin/gh/fxdawnn/9/head
2025-12-04T08:53:09.5381515Z  * [new branch]                gh/fxdawnn/9/orig       -> origin/gh/fxdawnn/9/orig
2025-12-04T08:53:09.5381584Z  * [new branch]                gh/galv/1/base          -> origin/gh/galv/1/base
2025-12-04T08:53:09.5381649Z  * [new branch]                gh/galv/1/head          -> origin/gh/galv/1/head
2025-12-04T08:53:09.5381712Z  * [new branch]                gh/galv/1/orig          -> origin/gh/galv/1/orig
2025-12-04T08:53:09.5381776Z  * [new branch]                gh/galv/2/base          -> origin/gh/galv/2/base
2025-12-04T08:53:09.5381841Z  * [new branch]                gh/galv/2/head          -> origin/gh/galv/2/head
2025-12-04T08:53:09.5381904Z  * [new branch]                gh/galv/2/orig          -> origin/gh/galv/2/orig
2025-12-04T08:53:09.5381967Z  * [new branch]                gh/galv/3/base          -> origin/gh/galv/3/base
2025-12-04T08:53:09.5382030Z  * [new branch]                gh/galv/3/head          -> origin/gh/galv/3/head
2025-12-04T08:53:09.5382092Z  * [new branch]                gh/galv/3/orig          -> origin/gh/galv/3/orig
2025-12-04T08:53:09.5382171Z  * [new branch]                gh/guangyey/134/base    -> origin/gh/guangyey/134/base
2025-12-04T08:53:09.5382247Z  * [new branch]                gh/guangyey/134/head    -> origin/gh/guangyey/134/head
2025-12-04T08:53:09.5382320Z  * [new branch]                gh/guangyey/134/orig    -> origin/gh/guangyey/134/orig
2025-12-04T08:53:09.5382392Z  * [new branch]                gh/guangyey/163/base    -> origin/gh/guangyey/163/base
2025-12-04T08:53:09.5382501Z  * [new branch]                gh/guangyey/163/head    -> origin/gh/guangyey/163/head
2025-12-04T08:53:09.5382615Z  * [new branch]                gh/guangyey/163/orig    -> origin/gh/guangyey/163/orig
2025-12-04T08:53:09.5382686Z  * [new branch]                gh/guangyey/168/base    -> origin/gh/guangyey/168/base
2025-12-04T08:53:09.5382758Z  * [new branch]                gh/guangyey/168/head    -> origin/gh/guangyey/168/head
2025-12-04T08:53:09.5382829Z  * [new branch]                gh/guangyey/168/orig    -> origin/gh/guangyey/168/orig
2025-12-04T08:53:09.5382900Z  * [new branch]                gh/guangyey/169/base    -> origin/gh/guangyey/169/base
2025-12-04T08:53:09.5382971Z  * [new branch]                gh/guangyey/169/head    -> origin/gh/guangyey/169/head
2025-12-04T08:53:09.5383043Z  * [new branch]                gh/guangyey/169/orig    -> origin/gh/guangyey/169/orig
2025-12-04T08:53:09.5383114Z  * [new branch]                gh/guangyey/170/base    -> origin/gh/guangyey/170/base
2025-12-04T08:53:09.5383188Z  * [new branch]                gh/guangyey/170/head    -> origin/gh/guangyey/170/head
2025-12-04T08:53:09.5383262Z  * [new branch]                gh/guangyey/170/orig    -> origin/gh/guangyey/170/orig
2025-12-04T08:53:09.5383332Z  * [new branch]                gh/guangyey/171/base    -> origin/gh/guangyey/171/base
2025-12-04T08:53:09.5383404Z  * [new branch]                gh/guangyey/171/head    -> origin/gh/guangyey/171/head
2025-12-04T08:53:09.5383477Z  * [new branch]                gh/guangyey/171/orig    -> origin/gh/guangyey/171/orig
2025-12-04T08:53:09.5383547Z  * [new branch]                gh/guangyey/178/base    -> origin/gh/guangyey/178/base
2025-12-04T08:53:09.5383618Z  * [new branch]                gh/guangyey/178/head    -> origin/gh/guangyey/178/head
2025-12-04T08:53:09.5383690Z  * [new branch]                gh/guangyey/178/orig    -> origin/gh/guangyey/178/orig
2025-12-04T08:53:09.5383762Z  * [new branch]                gh/guangyey/182/base    -> origin/gh/guangyey/182/base
2025-12-04T08:53:09.5383834Z  * [new branch]                gh/guangyey/182/head    -> origin/gh/guangyey/182/head
2025-12-04T08:53:09.5383908Z  * [new branch]                gh/guangyey/182/orig    -> origin/gh/guangyey/182/orig
2025-12-04T08:53:09.5383979Z  * [new branch]                gh/guangyey/183/base    -> origin/gh/guangyey/183/base
2025-12-04T08:53:09.5384051Z  * [new branch]                gh/guangyey/183/head    -> origin/gh/guangyey/183/head
2025-12-04T08:53:09.5384122Z  * [new branch]                gh/guangyey/183/orig    -> origin/gh/guangyey/183/orig
2025-12-04T08:53:09.5384193Z  * [new branch]                gh/guangyey/185/base    -> origin/gh/guangyey/185/base
2025-12-04T08:53:09.5384266Z  * [new branch]                gh/guangyey/185/head    -> origin/gh/guangyey/185/head
2025-12-04T08:53:09.5384337Z  * [new branch]                gh/guangyey/185/orig    -> origin/gh/guangyey/185/orig
2025-12-04T08:53:09.5384408Z  * [new branch]                gh/guangyey/186/base    -> origin/gh/guangyey/186/base
2025-12-04T08:53:09.5384482Z  * [new branch]                gh/guangyey/186/head    -> origin/gh/guangyey/186/head
2025-12-04T08:53:09.5384554Z  * [new branch]                gh/guangyey/186/orig    -> origin/gh/guangyey/186/orig
2025-12-04T08:53:09.5384625Z  * [new branch]                gh/guangyey/187/base    -> origin/gh/guangyey/187/base
2025-12-04T08:53:09.5384698Z  * [new branch]                gh/guangyey/187/head    -> origin/gh/guangyey/187/head
2025-12-04T08:53:09.5384768Z  * [new branch]                gh/guangyey/187/orig    -> origin/gh/guangyey/187/orig
2025-12-04T08:53:09.5384839Z  * [new branch]                gh/guangyey/188/base    -> origin/gh/guangyey/188/base
2025-12-04T08:53:09.5384912Z  * [new branch]                gh/guangyey/188/head    -> origin/gh/guangyey/188/head
2025-12-04T08:53:09.5384985Z  * [new branch]                gh/guangyey/188/orig    -> origin/gh/guangyey/188/orig
2025-12-04T08:53:09.5385055Z  * [new branch]                gh/guangyey/190/base    -> origin/gh/guangyey/190/base
2025-12-04T08:53:09.5385199Z  * [new branch]                gh/guangyey/190/head    -> origin/gh/guangyey/190/head
2025-12-04T08:53:09.5385301Z  * [new branch]                gh/guangyey/190/orig    -> origin/gh/guangyey/190/orig
2025-12-04T08:53:09.5385372Z  * [new branch]                gh/guangyey/208/base    -> origin/gh/guangyey/208/base
2025-12-04T08:53:09.5385444Z  * [new branch]                gh/guangyey/208/head    -> origin/gh/guangyey/208/head
2025-12-04T08:53:09.5385515Z  * [new branch]                gh/guangyey/208/orig    -> origin/gh/guangyey/208/orig
2025-12-04T08:53:09.5385587Z  * [new branch]                gh/guangyey/228/base    -> origin/gh/guangyey/228/base
2025-12-04T08:53:09.5385659Z  * [new branch]                gh/guangyey/228/head    -> origin/gh/guangyey/228/head
2025-12-04T08:53:09.5385730Z  * [new branch]                gh/guangyey/228/orig    -> origin/gh/guangyey/228/orig
2025-12-04T08:53:09.5385804Z  * [new branch]                gh/guangyey/230/base    -> origin/gh/guangyey/230/base
2025-12-04T08:53:09.5385877Z  * [new branch]                gh/guangyey/230/head    -> origin/gh/guangyey/230/head
2025-12-04T08:53:09.5385950Z  * [new branch]                gh/guangyey/230/orig    -> origin/gh/guangyey/230/orig
2025-12-04T08:53:09.5386023Z  * [new branch]                gh/guangyey/231/base    -> origin/gh/guangyey/231/base
2025-12-04T08:53:09.5386094Z  * [new branch]                gh/guangyey/231/head    -> origin/gh/guangyey/231/head
2025-12-04T08:53:09.5386165Z  * [new branch]                gh/guangyey/231/orig    -> origin/gh/guangyey/231/orig
2025-12-04T08:53:09.5386238Z  * [new branch]                gh/guangyey/232/base    -> origin/gh/guangyey/232/base
2025-12-04T08:53:09.5386309Z  * [new branch]                gh/guangyey/232/head    -> origin/gh/guangyey/232/head
2025-12-04T08:53:09.5386380Z  * [new branch]                gh/guangyey/232/orig    -> origin/gh/guangyey/232/orig
2025-12-04T08:53:09.5386455Z  * [new branch]                gh/guangyey/233/base    -> origin/gh/guangyey/233/base
2025-12-04T08:53:09.5386526Z  * [new branch]                gh/guangyey/233/head    -> origin/gh/guangyey/233/head
2025-12-04T08:53:09.5386597Z  * [new branch]                gh/guangyey/233/orig    -> origin/gh/guangyey/233/orig
2025-12-04T08:53:09.5386669Z  * [new branch]                gh/guangyey/234/base    -> origin/gh/guangyey/234/base
2025-12-04T08:53:09.5386740Z  * [new branch]                gh/guangyey/234/head    -> origin/gh/guangyey/234/head
2025-12-04T08:53:09.5386811Z  * [new branch]                gh/guangyey/234/orig    -> origin/gh/guangyey/234/orig
2025-12-04T08:53:09.5386884Z  * [new branch]                gh/guangyey/235/base    -> origin/gh/guangyey/235/base
2025-12-04T08:53:09.5386954Z  * [new branch]                gh/guangyey/235/head    -> origin/gh/guangyey/235/head
2025-12-04T08:53:09.5387027Z  * [new branch]                gh/guangyey/235/orig    -> origin/gh/guangyey/235/orig
2025-12-04T08:53:09.5387099Z  * [new branch]                gh/guangyey/236/base    -> origin/gh/guangyey/236/base
2025-12-04T08:53:09.5387170Z  * [new branch]                gh/guangyey/236/head    -> origin/gh/guangyey/236/head
2025-12-04T08:53:09.5387247Z  * [new branch]                gh/guangyey/236/orig    -> origin/gh/guangyey/236/orig
2025-12-04T08:53:09.5387318Z  * [new branch]                gh/guangyey/237/base    -> origin/gh/guangyey/237/base
2025-12-04T08:53:09.5387389Z  * [new branch]                gh/guangyey/237/head    -> origin/gh/guangyey/237/head
2025-12-04T08:53:09.5387461Z  * [new branch]                gh/guangyey/237/orig    -> origin/gh/guangyey/237/orig
2025-12-04T08:53:09.5387532Z  * [new branch]                gh/guangyey/238/base    -> origin/gh/guangyey/238/base
2025-12-04T08:53:09.5387602Z  * [new branch]                gh/guangyey/238/head    -> origin/gh/guangyey/238/head
2025-12-04T08:53:09.5387674Z  * [new branch]                gh/guangyey/239/base    -> origin/gh/guangyey/239/base
2025-12-04T08:53:09.5387772Z  * [new branch]                gh/guangyey/239/head    -> origin/gh/guangyey/239/head
2025-12-04T08:53:09.5387843Z  * [new branch]                gh/guangyey/239/orig    -> origin/gh/guangyey/239/orig
2025-12-04T08:53:09.5387968Z  * [new branch]                gh/guangyey/240/base    -> origin/gh/guangyey/240/base
2025-12-04T08:53:09.5388039Z  * [new branch]                gh/guangyey/240/head    -> origin/gh/guangyey/240/head
2025-12-04T08:53:09.5388110Z  * [new branch]                gh/guangyey/240/orig    -> origin/gh/guangyey/240/orig
2025-12-04T08:53:09.5388182Z  * [new branch]                gh/guangyey/241/base    -> origin/gh/guangyey/241/base
2025-12-04T08:53:09.5388253Z  * [new branch]                gh/guangyey/241/head    -> origin/gh/guangyey/241/head
2025-12-04T08:53:09.5388324Z  * [new branch]                gh/guangyey/241/orig    -> origin/gh/guangyey/241/orig
2025-12-04T08:53:09.5388396Z  * [new branch]                gh/guangyey/242/base    -> origin/gh/guangyey/242/base
2025-12-04T08:53:09.5388468Z  * [new branch]                gh/guangyey/242/head    -> origin/gh/guangyey/242/head
2025-12-04T08:53:09.5388539Z  * [new branch]                gh/guangyey/242/orig    -> origin/gh/guangyey/242/orig
2025-12-04T08:53:09.5388613Z  * [new branch]                gh/guangyey/243/base    -> origin/gh/guangyey/243/base
2025-12-04T08:53:09.5388684Z  * [new branch]                gh/guangyey/243/head    -> origin/gh/guangyey/243/head
2025-12-04T08:53:09.5388757Z  * [new branch]                gh/guangyey/243/orig    -> origin/gh/guangyey/243/orig
2025-12-04T08:53:09.5388828Z  * [new branch]                gh/guangyey/244/base    -> origin/gh/guangyey/244/base
2025-12-04T08:53:09.5388899Z  * [new branch]                gh/guangyey/244/head    -> origin/gh/guangyey/244/head
2025-12-04T08:53:09.5388972Z  * [new branch]                gh/guangyey/244/orig    -> origin/gh/guangyey/244/orig
2025-12-04T08:53:09.5389043Z  * [new branch]                gh/guangyey/245/base    -> origin/gh/guangyey/245/base
2025-12-04T08:53:09.5389115Z  * [new branch]                gh/guangyey/245/head    -> origin/gh/guangyey/245/head
2025-12-04T08:53:09.5389189Z  * [new branch]                gh/guangyey/245/orig    -> origin/gh/guangyey/245/orig
2025-12-04T08:53:09.5389259Z  * [new branch]                gh/guangyey/246/base    -> origin/gh/guangyey/246/base
2025-12-04T08:53:09.5389330Z  * [new branch]                gh/guangyey/246/head    -> origin/gh/guangyey/246/head
2025-12-04T08:53:09.5389402Z  * [new branch]                gh/guangyey/246/orig    -> origin/gh/guangyey/246/orig
2025-12-04T08:53:09.5389473Z  * [new branch]                gh/guangyey/247/base    -> origin/gh/guangyey/247/base
2025-12-04T08:53:09.5389543Z  * [new branch]                gh/guangyey/247/head    -> origin/gh/guangyey/247/head
2025-12-04T08:53:09.5389616Z  * [new branch]                gh/guangyey/247/orig    -> origin/gh/guangyey/247/orig
2025-12-04T08:53:09.5389687Z  * [new branch]                gh/guangyey/248/base    -> origin/gh/guangyey/248/base
2025-12-04T08:53:09.5389759Z  * [new branch]                gh/guangyey/248/head    -> origin/gh/guangyey/248/head
2025-12-04T08:53:09.5389833Z  * [new branch]                gh/guangyey/248/orig    -> origin/gh/guangyey/248/orig
2025-12-04T08:53:09.5389904Z  * [new branch]                gh/guangyey/249/base    -> origin/gh/guangyey/249/base
2025-12-04T08:53:09.5389974Z  * [new branch]                gh/guangyey/249/head    -> origin/gh/guangyey/249/head
2025-12-04T08:53:09.5390047Z  * [new branch]                gh/guangyey/249/orig    -> origin/gh/guangyey/249/orig
2025-12-04T08:53:09.5390118Z  * [new branch]                gh/guangyey/250/base    -> origin/gh/guangyey/250/base
2025-12-04T08:53:09.5390191Z  * [new branch]                gh/guangyey/250/head    -> origin/gh/guangyey/250/head
2025-12-04T08:53:09.5390262Z  * [new branch]                gh/guangyey/250/orig    -> origin/gh/guangyey/250/orig
2025-12-04T08:53:09.5390332Z  * [new branch]                gh/guangyey/251/base    -> origin/gh/guangyey/251/base
2025-12-04T08:53:09.5390465Z  * [new branch]                gh/guangyey/251/head    -> origin/gh/guangyey/251/head
2025-12-04T08:53:09.5390576Z  * [new branch]                gh/guangyey/251/orig    -> origin/gh/guangyey/251/orig
2025-12-04T08:53:09.5390647Z  * [new branch]                gh/guangyey/252/base    -> origin/gh/guangyey/252/base
2025-12-04T08:53:09.5390719Z  * [new branch]                gh/guangyey/252/head    -> origin/gh/guangyey/252/head
2025-12-04T08:53:09.5390790Z  * [new branch]                gh/guangyey/252/orig    -> origin/gh/guangyey/252/orig
2025-12-04T08:53:09.5390861Z  * [new branch]                gh/guangyey/253/base    -> origin/gh/guangyey/253/base
2025-12-04T08:53:09.5390932Z  * [new branch]                gh/guangyey/253/head    -> origin/gh/guangyey/253/head
2025-12-04T08:53:09.5391002Z  * [new branch]                gh/guangyey/253/orig    -> origin/gh/guangyey/253/orig
2025-12-04T08:53:09.5391073Z  * [new branch]                gh/guangyey/254/base    -> origin/gh/guangyey/254/base
2025-12-04T08:53:09.5391148Z  * [new branch]                gh/guangyey/254/head    -> origin/gh/guangyey/254/head
2025-12-04T08:53:09.5391220Z  * [new branch]                gh/guangyey/254/orig    -> origin/gh/guangyey/254/orig
2025-12-04T08:53:09.5391290Z  * [new branch]                gh/guangyey/255/base    -> origin/gh/guangyey/255/base
2025-12-04T08:53:09.5391363Z  * [new branch]                gh/guangyey/255/head    -> origin/gh/guangyey/255/head
2025-12-04T08:53:09.5391433Z  * [new branch]                gh/guangyey/255/orig    -> origin/gh/guangyey/255/orig
2025-12-04T08:53:09.5391532Z  * [new branch]                gh/guilhermeleobas/107/base -> origin/gh/guilhermeleobas/107/base
2025-12-04T08:53:09.5391627Z  * [new branch]                gh/guilhermeleobas/107/head -> origin/gh/guilhermeleobas/107/head
2025-12-04T08:53:09.5391717Z  * [new branch]                gh/guilhermeleobas/107/orig -> origin/gh/guilhermeleobas/107/orig
2025-12-04T08:53:09.5391809Z  * [new branch]                gh/guilhermeleobas/108/base -> origin/gh/guilhermeleobas/108/base
2025-12-04T08:53:09.5391897Z  * [new branch]                gh/guilhermeleobas/108/head -> origin/gh/guilhermeleobas/108/head
2025-12-04T08:53:09.5391986Z  * [new branch]                gh/guilhermeleobas/108/orig -> origin/gh/guilhermeleobas/108/orig
2025-12-04T08:53:09.5392075Z  * [new branch]                gh/guilhermeleobas/150/base -> origin/gh/guilhermeleobas/150/base
2025-12-04T08:53:09.5392163Z  * [new branch]                gh/guilhermeleobas/150/head -> origin/gh/guilhermeleobas/150/head
2025-12-04T08:53:09.5392252Z  * [new branch]                gh/guilhermeleobas/150/orig -> origin/gh/guilhermeleobas/150/orig
2025-12-04T08:53:09.5392340Z  * [new branch]                gh/guilhermeleobas/168/base -> origin/gh/guilhermeleobas/168/base
2025-12-04T08:53:09.5392427Z  * [new branch]                gh/guilhermeleobas/168/head -> origin/gh/guilhermeleobas/168/head
2025-12-04T08:53:09.5392514Z  * [new branch]                gh/guilhermeleobas/168/orig -> origin/gh/guilhermeleobas/168/orig
2025-12-04T08:53:09.5392607Z  * [new branch]                gh/guilhermeleobas/169/base -> origin/gh/guilhermeleobas/169/base
2025-12-04T08:53:09.5392696Z  * [new branch]                gh/guilhermeleobas/169/head -> origin/gh/guilhermeleobas/169/head
2025-12-04T08:53:09.5392783Z  * [new branch]                gh/guilhermeleobas/169/orig -> origin/gh/guilhermeleobas/169/orig
2025-12-04T08:53:09.5392872Z  * [new branch]                gh/guilhermeleobas/170/base -> origin/gh/guilhermeleobas/170/base
2025-12-04T08:53:09.5392960Z  * [new branch]                gh/guilhermeleobas/170/head -> origin/gh/guilhermeleobas/170/head
2025-12-04T08:53:09.5393048Z  * [new branch]                gh/guilhermeleobas/170/orig -> origin/gh/guilhermeleobas/170/orig
2025-12-04T08:53:09.5393137Z  * [new branch]                gh/guilhermeleobas/171/base -> origin/gh/guilhermeleobas/171/base
2025-12-04T08:53:09.5393225Z  * [new branch]                gh/guilhermeleobas/171/head -> origin/gh/guilhermeleobas/171/head
2025-12-04T08:53:09.5393356Z  * [new branch]                gh/guilhermeleobas/171/orig -> origin/gh/guilhermeleobas/171/orig
2025-12-04T08:53:09.5393470Z  * [new branch]                gh/guilhermeleobas/173/base -> origin/gh/guilhermeleobas/173/base
2025-12-04T08:53:09.5393558Z  * [new branch]                gh/guilhermeleobas/173/head -> origin/gh/guilhermeleobas/173/head
2025-12-04T08:53:09.5393648Z  * [new branch]                gh/guilhermeleobas/173/orig -> origin/gh/guilhermeleobas/173/orig
2025-12-04T08:53:09.5393735Z  * [new branch]                gh/guilhermeleobas/193/base -> origin/gh/guilhermeleobas/193/base
2025-12-04T08:53:09.5393822Z  * [new branch]                gh/guilhermeleobas/193/head -> origin/gh/guilhermeleobas/193/head
2025-12-04T08:53:09.5393911Z  * [new branch]                gh/guilhermeleobas/193/orig -> origin/gh/guilhermeleobas/193/orig
2025-12-04T08:53:09.5393999Z  * [new branch]                gh/guilhermeleobas/204/base -> origin/gh/guilhermeleobas/204/base
2025-12-04T08:53:09.5394090Z  * [new branch]                gh/guilhermeleobas/204/head -> origin/gh/guilhermeleobas/204/head
2025-12-04T08:53:09.5394182Z  * [new branch]                gh/guilhermeleobas/204/orig -> origin/gh/guilhermeleobas/204/orig
2025-12-04T08:53:09.5394269Z  * [new branch]                gh/guilhermeleobas/211/base -> origin/gh/guilhermeleobas/211/base
2025-12-04T08:53:09.5394356Z  * [new branch]                gh/guilhermeleobas/211/head -> origin/gh/guilhermeleobas/211/head
2025-12-04T08:53:09.5394445Z  * [new branch]                gh/guilhermeleobas/211/orig -> origin/gh/guilhermeleobas/211/orig
2025-12-04T08:53:09.5394532Z  * [new branch]                gh/guilhermeleobas/226/base -> origin/gh/guilhermeleobas/226/base
2025-12-04T08:53:09.5394619Z  * [new branch]                gh/guilhermeleobas/226/head -> origin/gh/guilhermeleobas/226/head
2025-12-04T08:53:09.5394708Z  * [new branch]                gh/guilhermeleobas/226/orig -> origin/gh/guilhermeleobas/226/orig
2025-12-04T08:53:09.5394797Z  * [new branch]                gh/guilhermeleobas/236/base -> origin/gh/guilhermeleobas/236/base
2025-12-04T08:53:09.5394888Z  * [new branch]                gh/guilhermeleobas/236/head -> origin/gh/guilhermeleobas/236/head
2025-12-04T08:53:09.5394975Z  * [new branch]                gh/guilhermeleobas/236/orig -> origin/gh/guilhermeleobas/236/orig
2025-12-04T08:53:09.5395062Z  * [new branch]                gh/guilhermeleobas/247/base -> origin/gh/guilhermeleobas/247/base
2025-12-04T08:53:09.5395150Z  * [new branch]                gh/guilhermeleobas/247/head -> origin/gh/guilhermeleobas/247/head
2025-12-04T08:53:09.5395237Z  * [new branch]                gh/guilhermeleobas/247/orig -> origin/gh/guilhermeleobas/247/orig
2025-12-04T08:53:09.5395325Z  * [new branch]                gh/guilhermeleobas/248/base -> origin/gh/guilhermeleobas/248/base
2025-12-04T08:53:09.5395413Z  * [new branch]                gh/guilhermeleobas/248/head -> origin/gh/guilhermeleobas/248/head
2025-12-04T08:53:09.5395503Z  * [new branch]                gh/guilhermeleobas/248/orig -> origin/gh/guilhermeleobas/248/orig
2025-12-04T08:53:09.5395592Z  * [new branch]                gh/guilhermeleobas/250/base -> origin/gh/guilhermeleobas/250/base
2025-12-04T08:53:09.5395683Z  * [new branch]                gh/guilhermeleobas/250/head -> origin/gh/guilhermeleobas/250/head
2025-12-04T08:53:09.5395770Z  * [new branch]                gh/guilhermeleobas/250/orig -> origin/gh/guilhermeleobas/250/orig
2025-12-04T08:53:09.5395858Z  * [new branch]                gh/guilhermeleobas/253/base -> origin/gh/guilhermeleobas/253/base
2025-12-04T08:53:09.5395947Z  * [new branch]                gh/guilhermeleobas/253/head -> origin/gh/guilhermeleobas/253/head
2025-12-04T08:53:09.5396034Z  * [new branch]                gh/guilhermeleobas/253/orig -> origin/gh/guilhermeleobas/253/orig
2025-12-04T08:53:09.5396121Z  * [new branch]                gh/guilhermeleobas/254/base -> origin/gh/guilhermeleobas/254/base
2025-12-04T08:53:09.5396240Z  * [new branch]                gh/guilhermeleobas/254/head -> origin/gh/guilhermeleobas/254/head
2025-12-04T08:53:09.5396347Z  * [new branch]                gh/guilhermeleobas/254/orig -> origin/gh/guilhermeleobas/254/orig
2025-12-04T08:53:09.5396436Z  * [new branch]                gh/guilhermeleobas/255/base -> origin/gh/guilhermeleobas/255/base
2025-12-04T08:53:09.5396524Z  * [new branch]                gh/guilhermeleobas/255/head -> origin/gh/guilhermeleobas/255/head
2025-12-04T08:53:09.5396612Z  * [new branch]                gh/guilhermeleobas/255/orig -> origin/gh/guilhermeleobas/255/orig
2025-12-04T08:53:09.5396701Z  * [new branch]                gh/guilhermeleobas/256/base -> origin/gh/guilhermeleobas/256/base
2025-12-04T08:53:09.5396788Z  * [new branch]                gh/guilhermeleobas/256/head -> origin/gh/guilhermeleobas/256/head
2025-12-04T08:53:09.5396875Z  * [new branch]                gh/guilhermeleobas/256/orig -> origin/gh/guilhermeleobas/256/orig
2025-12-04T08:53:09.5396968Z  * [new branch]                gh/guilhermeleobas/257/base -> origin/gh/guilhermeleobas/257/base
2025-12-04T08:53:09.5397056Z  * [new branch]                gh/guilhermeleobas/257/head -> origin/gh/guilhermeleobas/257/head
2025-12-04T08:53:09.5397143Z  * [new branch]                gh/guilhermeleobas/257/orig -> origin/gh/guilhermeleobas/257/orig
2025-12-04T08:53:09.5397232Z  * [new branch]                gh/guilhermeleobas/258/base -> origin/gh/guilhermeleobas/258/base
2025-12-04T08:53:09.5397319Z  * [new branch]                gh/guilhermeleobas/258/head -> origin/gh/guilhermeleobas/258/head
2025-12-04T08:53:09.5397405Z  * [new branch]                gh/guilhermeleobas/258/orig -> origin/gh/guilhermeleobas/258/orig
2025-12-04T08:53:09.5397495Z  * [new branch]                gh/guilhermeleobas/259/base -> origin/gh/guilhermeleobas/259/base
2025-12-04T08:53:09.5397582Z  * [new branch]                gh/guilhermeleobas/259/head -> origin/gh/guilhermeleobas/259/head
2025-12-04T08:53:09.5397671Z  * [new branch]                gh/guilhermeleobas/259/orig -> origin/gh/guilhermeleobas/259/orig
2025-12-04T08:53:09.5397761Z  * [new branch]                gh/guilhermeleobas/260/base -> origin/gh/guilhermeleobas/260/base
2025-12-04T08:53:09.5397849Z  * [new branch]                gh/guilhermeleobas/260/head -> origin/gh/guilhermeleobas/260/head
2025-12-04T08:53:09.5397938Z  * [new branch]                gh/guilhermeleobas/260/orig -> origin/gh/guilhermeleobas/260/orig
2025-12-04T08:53:09.5398026Z  * [new branch]                gh/guilhermeleobas/261/base -> origin/gh/guilhermeleobas/261/base
2025-12-04T08:53:09.5398112Z  * [new branch]                gh/guilhermeleobas/261/head -> origin/gh/guilhermeleobas/261/head
2025-12-04T08:53:09.5398202Z  * [new branch]                gh/guilhermeleobas/261/orig -> origin/gh/guilhermeleobas/261/orig
2025-12-04T08:53:09.5398289Z  * [new branch]                gh/guilhermeleobas/262/base -> origin/gh/guilhermeleobas/262/base
2025-12-04T08:53:09.5398376Z  * [new branch]                gh/guilhermeleobas/262/head -> origin/gh/guilhermeleobas/262/head
2025-12-04T08:53:09.5398466Z  * [new branch]                gh/guilhermeleobas/262/orig -> origin/gh/guilhermeleobas/262/orig
2025-12-04T08:53:09.5398556Z  * [new branch]                gh/guilhermeleobas/263/base -> origin/gh/guilhermeleobas/263/base
2025-12-04T08:53:09.5398643Z  * [new branch]                gh/guilhermeleobas/263/head -> origin/gh/guilhermeleobas/263/head
2025-12-04T08:53:09.5398732Z  * [new branch]                gh/guilhermeleobas/263/orig -> origin/gh/guilhermeleobas/263/orig
2025-12-04T08:53:09.5398821Z  * [new branch]                gh/guilhermeleobas/264/base -> origin/gh/guilhermeleobas/264/base
2025-12-04T08:53:09.5398908Z  * [new branch]                gh/guilhermeleobas/264/head -> origin/gh/guilhermeleobas/264/head
2025-12-04T08:53:09.5398998Z  * [new branch]                gh/guilhermeleobas/264/orig -> origin/gh/guilhermeleobas/264/orig
2025-12-04T08:53:09.5399085Z  * [new branch]                gh/guilhermeleobas/265/base -> origin/gh/guilhermeleobas/265/base
2025-12-04T08:53:09.5399195Z  * [new branch]                gh/guilhermeleobas/265/head -> origin/gh/guilhermeleobas/265/head
2025-12-04T08:53:09.5399307Z  * [new branch]                gh/guilhermeleobas/265/orig -> origin/gh/guilhermeleobas/265/orig
2025-12-04T08:53:09.5399395Z  * [new branch]                gh/guilhermeleobas/266/base -> origin/gh/guilhermeleobas/266/base
2025-12-04T08:53:09.5399484Z  * [new branch]                gh/guilhermeleobas/266/head -> origin/gh/guilhermeleobas/266/head
2025-12-04T08:53:09.5399570Z  * [new branch]                gh/guilhermeleobas/266/orig -> origin/gh/guilhermeleobas/266/orig
2025-12-04T08:53:09.5399658Z  * [new branch]                gh/guilhermeleobas/267/base -> origin/gh/guilhermeleobas/267/base
2025-12-04T08:53:09.5399747Z  * [new branch]                gh/guilhermeleobas/267/head -> origin/gh/guilhermeleobas/267/head
2025-12-04T08:53:09.5399834Z  * [new branch]                gh/guilhermeleobas/267/orig -> origin/gh/guilhermeleobas/267/orig
2025-12-04T08:53:09.5399918Z  * [new branch]                gh/hameerabbasi/1/base  -> origin/gh/hameerabbasi/1/base
2025-12-04T08:53:09.5399998Z  * [new branch]                gh/hameerabbasi/1/head  -> origin/gh/hameerabbasi/1/head
2025-12-04T08:53:09.5400074Z  * [new branch]                gh/hameerabbasi/2/base  -> origin/gh/hameerabbasi/2/base
2025-12-04T08:53:09.5400149Z  * [new branch]                gh/hameerabbasi/2/head  -> origin/gh/hameerabbasi/2/head
2025-12-04T08:53:09.5400227Z  * [new branch]                gh/hameerabbasi/2/orig  -> origin/gh/hameerabbasi/2/orig
2025-12-04T08:53:09.5400301Z  * [new branch]                gh/hameerabbasi/3/base  -> origin/gh/hameerabbasi/3/base
2025-12-04T08:53:09.5400377Z  * [new branch]                gh/hameerabbasi/3/head  -> origin/gh/hameerabbasi/3/head
2025-12-04T08:53:09.5400469Z  * [new branch]                gh/hameerabbasi/3/orig  -> origin/gh/hameerabbasi/3/orig
2025-12-04T08:53:09.5400545Z  * [new branch]                gh/hameerabbasi/4/base  -> origin/gh/hameerabbasi/4/base
2025-12-04T08:53:09.5400621Z  * [new branch]                gh/hameerabbasi/4/head  -> origin/gh/hameerabbasi/4/head
2025-12-04T08:53:09.5400698Z  * [new branch]                gh/hameerabbasi/4/orig  -> origin/gh/hameerabbasi/4/orig
2025-12-04T08:53:09.5400768Z  * [new branch]                gh/huydhn/1/next        -> origin/gh/huydhn/1/next
2025-12-04T08:53:09.5400837Z  * [new branch]                gh/huydhn/2/next        -> origin/gh/huydhn/2/next
2025-12-04T08:53:09.5400904Z  * [new branch]                gh/huydhn/3/next        -> origin/gh/huydhn/3/next
2025-12-04T08:53:09.5400970Z  * [new branch]                gh/huydhn/4/next        -> origin/gh/huydhn/4/next
2025-12-04T08:53:09.5401036Z  * [new branch]                gh/huydhn/5/next        -> origin/gh/huydhn/5/next
2025-12-04T08:53:09.5401102Z  * [new branch]                gh/huydhn/6/next        -> origin/gh/huydhn/6/next
2025-12-04T08:53:09.5401168Z  * [new branch]                gh/int3/97/base         -> origin/gh/int3/97/base
2025-12-04T08:53:09.5401237Z  * [new branch]                gh/int3/97/head         -> origin/gh/int3/97/head
2025-12-04T08:53:09.5401309Z  * [new branch]                gh/isuruf/101/base      -> origin/gh/isuruf/101/base
2025-12-04T08:53:09.5401379Z  * [new branch]                gh/isuruf/101/head      -> origin/gh/isuruf/101/head
2025-12-04T08:53:09.5401449Z  * [new branch]                gh/isuruf/146/base      -> origin/gh/isuruf/146/base
2025-12-04T08:53:09.5401519Z  * [new branch]                gh/isuruf/146/head      -> origin/gh/isuruf/146/head
2025-12-04T08:53:09.5401587Z  * [new branch]                gh/isuruf/146/orig      -> origin/gh/isuruf/146/orig
2025-12-04T08:53:09.5401656Z  * [new branch]                gh/isuruf/158/base      -> origin/gh/isuruf/158/base
2025-12-04T08:53:09.5401723Z  * [new branch]                gh/isuruf/158/head      -> origin/gh/isuruf/158/head
2025-12-04T08:53:09.5401790Z  * [new branch]                gh/isuruf/159/base      -> origin/gh/isuruf/159/base
2025-12-04T08:53:09.5401905Z  * [new branch]                gh/isuruf/159/head      -> origin/gh/isuruf/159/head
2025-12-04T08:53:09.5402012Z  * [new branch]                gh/isuruf/160/base      -> origin/gh/isuruf/160/base
2025-12-04T08:53:09.5402079Z  * [new branch]                gh/isuruf/160/head      -> origin/gh/isuruf/160/head
2025-12-04T08:53:09.5402148Z  * [new branch]                gh/isuruf/160/orig      -> origin/gh/isuruf/160/orig
2025-12-04T08:53:09.5402216Z  * [new branch]                gh/isuruf/81/base       -> origin/gh/isuruf/81/base
2025-12-04T08:53:09.5402284Z  * [new branch]                gh/isuruf/81/head       -> origin/gh/isuruf/81/head
2025-12-04T08:53:09.5402352Z  * [new branch]                gh/isuruf/81/orig       -> origin/gh/isuruf/81/orig
2025-12-04T08:53:09.5402426Z  * [new branch]                gh/jamesjwu/176/base    -> origin/gh/jamesjwu/176/base
2025-12-04T08:53:09.5402499Z  * [new branch]                gh/jamesjwu/176/head    -> origin/gh/jamesjwu/176/head
2025-12-04T08:53:09.5402576Z  * [new branch]                gh/jamesjwu/176/orig    -> origin/gh/jamesjwu/176/orig
2025-12-04T08:53:09.5402649Z  * [new branch]                gh/jamesjwu/187/base    -> origin/gh/jamesjwu/187/base
2025-12-04T08:53:09.5402722Z  * [new branch]                gh/jamesjwu/187/head    -> origin/gh/jamesjwu/187/head
2025-12-04T08:53:09.5402792Z  * [new branch]                gh/jamesjwu/187/orig    -> origin/gh/jamesjwu/187/orig
2025-12-04T08:53:09.5402863Z  * [new branch]                gh/jamesjwu/196/base    -> origin/gh/jamesjwu/196/base
2025-12-04T08:53:09.5402937Z  * [new branch]                gh/jamesjwu/196/head    -> origin/gh/jamesjwu/196/head
2025-12-04T08:53:09.5403008Z  * [new branch]                gh/jamesjwu/196/orig    -> origin/gh/jamesjwu/196/orig
2025-12-04T08:53:09.5403078Z  * [new branch]                gh/jamesjwu/198/base    -> origin/gh/jamesjwu/198/base
2025-12-04T08:53:09.5403150Z  * [new branch]                gh/jamesjwu/198/head    -> origin/gh/jamesjwu/198/head
2025-12-04T08:53:09.5403222Z  * [new branch]                gh/jamesjwu/198/orig    -> origin/gh/jamesjwu/198/orig
2025-12-04T08:53:09.5403294Z  * [new branch]                gh/jamesjwu/207/base    -> origin/gh/jamesjwu/207/base
2025-12-04T08:53:09.5403365Z  * [new branch]                gh/jamesjwu/207/head    -> origin/gh/jamesjwu/207/head
2025-12-04T08:53:09.5403435Z  * [new branch]                gh/jamesjwu/207/orig    -> origin/gh/jamesjwu/207/orig
2025-12-04T08:53:09.5403505Z  * [new branch]                gh/jamesjwu/208/base    -> origin/gh/jamesjwu/208/base
2025-12-04T08:53:09.5403577Z  * [new branch]                gh/jamesjwu/208/head    -> origin/gh/jamesjwu/208/head
2025-12-04T08:53:09.5403647Z  * [new branch]                gh/jamesjwu/208/orig    -> origin/gh/jamesjwu/208/orig
2025-12-04T08:53:09.5403718Z  * [new branch]                gh/jamesjwu/52/base     -> origin/gh/jamesjwu/52/base
2025-12-04T08:53:09.5403791Z  * [new branch]                gh/jamesjwu/52/head     -> origin/gh/jamesjwu/52/head
2025-12-04T08:53:09.5403862Z  * [new branch]                gh/jamesjwu/53/base     -> origin/gh/jamesjwu/53/base
2025-12-04T08:53:09.5403933Z  * [new branch]                gh/jamesjwu/53/head     -> origin/gh/jamesjwu/53/head
2025-12-04T08:53:09.5404003Z  * [new branch]                gh/jamesjwu/54/base     -> origin/gh/jamesjwu/54/base
2025-12-04T08:53:09.5404072Z  * [new branch]                gh/jamesjwu/54/head     -> origin/gh/jamesjwu/54/head
2025-12-04T08:53:09.5404143Z  * [new branch]                gh/jamesjwu/55/base     -> origin/gh/jamesjwu/55/base
2025-12-04T08:53:09.5404212Z  * [new branch]                gh/jamesjwu/55/head     -> origin/gh/jamesjwu/55/head
2025-12-04T08:53:09.5404280Z  * [new branch]                gh/jamesjwu/56/base     -> origin/gh/jamesjwu/56/base
2025-12-04T08:53:09.5404350Z  * [new branch]                gh/jamesjwu/56/head     -> origin/gh/jamesjwu/56/head
2025-12-04T08:53:09.5404420Z  * [new branch]                gh/jamesjwu/57/base     -> origin/gh/jamesjwu/57/base
2025-12-04T08:53:09.5404518Z  * [new branch]                gh/jamesjwu/57/head     -> origin/gh/jamesjwu/57/head
2025-12-04T08:53:09.5404614Z  * [new branch]                gh/jamesjwu/58/base     -> origin/gh/jamesjwu/58/base
2025-12-04T08:53:09.5404683Z  * [new branch]                gh/jamesjwu/58/head     -> origin/gh/jamesjwu/58/head
2025-12-04T08:53:09.5404752Z  * [new branch]                gh/jamesjwu/59/base     -> origin/gh/jamesjwu/59/base
2025-12-04T08:53:09.5404823Z  * [new branch]                gh/jamesjwu/59/head     -> origin/gh/jamesjwu/59/head
2025-12-04T08:53:09.5404894Z  * [new branch]                gh/jamesjwu/60/base     -> origin/gh/jamesjwu/60/base
2025-12-04T08:53:09.5404963Z  * [new branch]                gh/jamesjwu/60/head     -> origin/gh/jamesjwu/60/head
2025-12-04T08:53:09.5405033Z  * [new branch]                gh/jamesjwu/61/base     -> origin/gh/jamesjwu/61/base
2025-12-04T08:53:09.5405102Z  * [new branch]                gh/jamesjwu/61/head     -> origin/gh/jamesjwu/61/head
2025-12-04T08:53:09.5405173Z  * [new branch]                gh/jamesjwu/62/base     -> origin/gh/jamesjwu/62/base
2025-12-04T08:53:09.5405245Z  * [new branch]                gh/jamesjwu/62/head     -> origin/gh/jamesjwu/62/head
2025-12-04T08:53:09.5405314Z  * [new branch]                gh/jamesjwu/63/base     -> origin/gh/jamesjwu/63/base
2025-12-04T08:53:09.5405384Z  * [new branch]                gh/jamesjwu/63/head     -> origin/gh/jamesjwu/63/head
2025-12-04T08:53:09.5405452Z  * [new branch]                gh/jamesjwu/64/base     -> origin/gh/jamesjwu/64/base
2025-12-04T08:53:09.5405522Z  * [new branch]                gh/jamesjwu/64/head     -> origin/gh/jamesjwu/64/head
2025-12-04T08:53:09.5405592Z  * [new branch]                gh/jamesjwu/65/base     -> origin/gh/jamesjwu/65/base
2025-12-04T08:53:09.5405661Z  * [new branch]                gh/jamesjwu/65/head     -> origin/gh/jamesjwu/65/head
2025-12-04T08:53:09.5405736Z  * [new branch]                gh/janeyx99/165/base    -> origin/gh/janeyx99/165/base
2025-12-04T08:53:09.5405807Z  * [new branch]                gh/janeyx99/165/head    -> origin/gh/janeyx99/165/head
2025-12-04T08:53:09.5405881Z  * [new branch]                gh/janeyx99/165/orig    -> origin/gh/janeyx99/165/orig
2025-12-04T08:53:09.5405953Z  * [new branch]                gh/janeyx99/201/base    -> origin/gh/janeyx99/201/base
2025-12-04T08:53:09.5406023Z  * [new branch]                gh/janeyx99/201/head    -> origin/gh/janeyx99/201/head
2025-12-04T08:53:09.5406092Z  * [new branch]                gh/janeyx99/201/orig    -> origin/gh/janeyx99/201/orig
2025-12-04T08:53:09.5406164Z  * [new branch]                gh/janeyx99/225/base    -> origin/gh/janeyx99/225/base
2025-12-04T08:53:09.5406232Z  * [new branch]                gh/janeyx99/225/head    -> origin/gh/janeyx99/225/head
2025-12-04T08:53:09.5406302Z  * [new branch]                gh/janeyx99/225/orig    -> origin/gh/janeyx99/225/orig
2025-12-04T08:53:09.5406375Z  * [new branch]                gh/janeyx99/299/base    -> origin/gh/janeyx99/299/base
2025-12-04T08:53:09.5406445Z  * [new branch]                gh/janeyx99/299/head    -> origin/gh/janeyx99/299/head
2025-12-04T08:53:09.5406519Z  * [new branch]                gh/janeyx99/299/orig    -> origin/gh/janeyx99/299/orig
2025-12-04T08:53:09.5406594Z  * [new branch]                gh/janeyx99/302/base    -> origin/gh/janeyx99/302/base
2025-12-04T08:53:09.5406663Z  * [new branch]                gh/janeyx99/302/head    -> origin/gh/janeyx99/302/head
2025-12-04T08:53:09.5406733Z  * [new branch]                gh/janeyx99/303/base    -> origin/gh/janeyx99/303/base
2025-12-04T08:53:09.5406805Z  * [new branch]                gh/janeyx99/303/head    -> origin/gh/janeyx99/303/head
2025-12-04T08:53:09.5406874Z  * [new branch]                gh/janeyx99/305/base    -> origin/gh/janeyx99/305/base
2025-12-04T08:53:09.5406945Z  * [new branch]                gh/janeyx99/305/head    -> origin/gh/janeyx99/305/head
2025-12-04T08:53:09.5407047Z  * [new branch]                gh/janeyx99/306/base    -> origin/gh/janeyx99/306/base
2025-12-04T08:53:09.5407117Z  * [new branch]                gh/janeyx99/306/head    -> origin/gh/janeyx99/306/head
2025-12-04T08:53:09.5407219Z  * [new branch]                gh/janeyx99/314/base    -> origin/gh/janeyx99/314/base
2025-12-04T08:53:09.5407290Z  * [new branch]                gh/janeyx99/314/head    -> origin/gh/janeyx99/314/head
2025-12-04T08:53:09.5407360Z  * [new branch]                gh/janeyx99/314/orig    -> origin/gh/janeyx99/314/orig
2025-12-04T08:53:09.5407432Z  * [new branch]                gh/janeyx99/315/base    -> origin/gh/janeyx99/315/base
2025-12-04T08:53:09.5407502Z  * [new branch]                gh/janeyx99/315/head    -> origin/gh/janeyx99/315/head
2025-12-04T08:53:09.5407571Z  * [new branch]                gh/janeyx99/315/orig    -> origin/gh/janeyx99/315/orig
2025-12-04T08:53:09.5407642Z  * [new branch]                gh/janeyx99/316/base    -> origin/gh/janeyx99/316/base
2025-12-04T08:53:09.5407714Z  * [new branch]                gh/janeyx99/316/head    -> origin/gh/janeyx99/316/head
2025-12-04T08:53:09.5407784Z  * [new branch]                gh/janeyx99/316/orig    -> origin/gh/janeyx99/316/orig
2025-12-04T08:53:09.5407856Z  * [new branch]                gh/janeyx99/317/base    -> origin/gh/janeyx99/317/base
2025-12-04T08:53:09.5407926Z  * [new branch]                gh/janeyx99/317/head    -> origin/gh/janeyx99/317/head
2025-12-04T08:53:09.5407998Z  * [new branch]                gh/janeyx99/317/orig    -> origin/gh/janeyx99/317/orig
2025-12-04T08:53:09.5408070Z  * [new branch]                gh/janeyx99/325/base    -> origin/gh/janeyx99/325/base
2025-12-04T08:53:09.5408140Z  * [new branch]                gh/janeyx99/325/head    -> origin/gh/janeyx99/325/head
2025-12-04T08:53:09.5408210Z  * [new branch]                gh/janeyx99/325/orig    -> origin/gh/janeyx99/325/orig
2025-12-04T08:53:09.5408281Z  * [new branch]                gh/janeyx99/327/base    -> origin/gh/janeyx99/327/base
2025-12-04T08:53:09.5408353Z  * [new branch]                gh/janeyx99/327/head    -> origin/gh/janeyx99/327/head
2025-12-04T08:53:09.5408423Z  * [new branch]                gh/janeyx99/327/orig    -> origin/gh/janeyx99/327/orig
2025-12-04T08:53:09.5408497Z  * [new branch]                gh/janeyx99/328/base    -> origin/gh/janeyx99/328/base
2025-12-04T08:53:09.5408567Z  * [new branch]                gh/janeyx99/328/head    -> origin/gh/janeyx99/328/head
2025-12-04T08:53:09.5408638Z  * [new branch]                gh/janeyx99/328/orig    -> origin/gh/janeyx99/328/orig
2025-12-04T08:53:09.5408707Z  * [new branch]                gh/janeyx99/329/base    -> origin/gh/janeyx99/329/base
2025-12-04T08:53:09.5408777Z  * [new branch]                gh/janeyx99/329/head    -> origin/gh/janeyx99/329/head
2025-12-04T08:53:09.5408849Z  * [new branch]                gh/janeyx99/329/orig    -> origin/gh/janeyx99/329/orig
2025-12-04T08:53:09.5408918Z  * [new branch]                gh/janeyx99/330/base    -> origin/gh/janeyx99/330/base
2025-12-04T08:53:09.5408989Z  * [new branch]                gh/janeyx99/330/head    -> origin/gh/janeyx99/330/head
2025-12-04T08:53:09.5409061Z  * [new branch]                gh/janeyx99/330/orig    -> origin/gh/janeyx99/330/orig
2025-12-04T08:53:09.5409131Z  * [new branch]                gh/janeyx99/331/base    -> origin/gh/janeyx99/331/base
2025-12-04T08:53:09.5409200Z  * [new branch]                gh/janeyx99/331/head    -> origin/gh/janeyx99/331/head
2025-12-04T08:53:09.5409272Z  * [new branch]                gh/janeyx99/331/orig    -> origin/gh/janeyx99/331/orig
2025-12-04T08:53:09.5409341Z  * [new branch]                gh/janeyx99/332/base    -> origin/gh/janeyx99/332/base
2025-12-04T08:53:09.5409411Z  * [new branch]                gh/janeyx99/332/head    -> origin/gh/janeyx99/332/head
2025-12-04T08:53:09.5409483Z  * [new branch]                gh/janeyx99/332/orig    -> origin/gh/janeyx99/332/orig
2025-12-04T08:53:09.5409553Z  * [new branch]                gh/janeyx99/333/base    -> origin/gh/janeyx99/333/base
2025-12-04T08:53:09.5409648Z  * [new branch]                gh/janeyx99/333/head    -> origin/gh/janeyx99/333/head
2025-12-04T08:53:09.5409741Z  * [new branch]                gh/janeyx99/333/orig    -> origin/gh/janeyx99/333/orig
2025-12-04T08:53:09.5409810Z  * [new branch]                gh/janeyx99/88/base     -> origin/gh/janeyx99/88/base
2025-12-04T08:53:09.5409879Z  * [new branch]                gh/janeyx99/88/head     -> origin/gh/janeyx99/88/head
2025-12-04T08:53:09.5409951Z  * [new branch]                gh/janeyx99/88/orig     -> origin/gh/janeyx99/88/orig
2025-12-04T08:53:09.5410021Z  * [new branch]                gh/jansel/360/base      -> origin/gh/jansel/360/base
2025-12-04T08:53:09.5410090Z  * [new branch]                gh/jansel/360/head      -> origin/gh/jansel/360/head
2025-12-04T08:53:09.5410158Z  * [new branch]                gh/jansel/451/base      -> origin/gh/jansel/451/base
2025-12-04T08:53:09.5410226Z  * [new branch]                gh/jansel/451/head      -> origin/gh/jansel/451/head
2025-12-04T08:53:09.5410298Z  * [new branch]                gh/jansel/451/orig      -> origin/gh/jansel/451/orig
2025-12-04T08:53:09.5410365Z  * [new branch]                gh/jansel/462/base      -> origin/gh/jansel/462/base
2025-12-04T08:53:09.5410465Z  * [new branch]                gh/jansel/462/head      -> origin/gh/jansel/462/head
2025-12-04T08:53:09.5410534Z  * [new branch]                gh/jansel/462/orig      -> origin/gh/jansel/462/orig
2025-12-04T08:53:09.5410602Z  * [new branch]                gh/jansel/533/base      -> origin/gh/jansel/533/base
2025-12-04T08:53:09.5410669Z  * [new branch]                gh/jansel/533/head      -> origin/gh/jansel/533/head
2025-12-04T08:53:09.5410737Z  * [new branch]                gh/jansel/533/orig      -> origin/gh/jansel/533/orig
2025-12-04T08:53:09.5410804Z  * [new branch]                gh/jansel/552/base      -> origin/gh/jansel/552/base
2025-12-04T08:53:09.5410873Z  * [new branch]                gh/jansel/552/head      -> origin/gh/jansel/552/head
2025-12-04T08:53:09.5410943Z  * [new branch]                gh/jansel/552/orig      -> origin/gh/jansel/552/orig
2025-12-04T08:53:09.5411011Z  * [new branch]                gh/jansel/553/base      -> origin/gh/jansel/553/base
2025-12-04T08:53:09.5411077Z  * [new branch]                gh/jansel/553/head      -> origin/gh/jansel/553/head
2025-12-04T08:53:09.5411147Z  * [new branch]                gh/jansel/553/orig      -> origin/gh/jansel/553/orig
2025-12-04T08:53:09.5411214Z  * [new branch]                gh/jansel/554/base      -> origin/gh/jansel/554/base
2025-12-04T08:53:09.5411281Z  * [new branch]                gh/jansel/554/head      -> origin/gh/jansel/554/head
2025-12-04T08:53:09.5411349Z  * [new branch]                gh/jansel/554/orig      -> origin/gh/jansel/554/orig
2025-12-04T08:53:09.5411416Z  * [new branch]                gh/jansel/555/base      -> origin/gh/jansel/555/base
2025-12-04T08:53:09.5411483Z  * [new branch]                gh/jansel/555/head      -> origin/gh/jansel/555/head
2025-12-04T08:53:09.5411553Z  * [new branch]                gh/jansel/555/orig      -> origin/gh/jansel/555/orig
2025-12-04T08:53:09.5411621Z  * [new branch]                gh/jansel/556/base      -> origin/gh/jansel/556/base
2025-12-04T08:53:09.5411690Z  * [new branch]                gh/jansel/556/head      -> origin/gh/jansel/556/head
2025-12-04T08:53:09.5411757Z  * [new branch]                gh/jansel/556/orig      -> origin/gh/jansel/556/orig
2025-12-04T08:53:09.5411825Z  * [new branch]                gh/jansel/557/base      -> origin/gh/jansel/557/base
2025-12-04T08:53:09.5411893Z  * [new branch]                gh/jansel/557/head      -> origin/gh/jansel/557/head
2025-12-04T08:53:09.5411960Z  * [new branch]                gh/jansel/557/orig      -> origin/gh/jansel/557/orig
2025-12-04T08:53:09.5412027Z  * [new branch]                gh/jansel/558/base      -> origin/gh/jansel/558/base
2025-12-04T08:53:09.5412094Z  * [new branch]                gh/jansel/558/head      -> origin/gh/jansel/558/head
2025-12-04T08:53:09.5412197Z  * [new branch]                gh/jansel/558/orig      -> origin/gh/jansel/558/orig
2025-12-04T08:53:09.5412263Z  * [new branch]                gh/jansel/559/base      -> origin/gh/jansel/559/base
2025-12-04T08:53:09.5412370Z  * [new branch]                gh/jansel/559/head      -> origin/gh/jansel/559/head
2025-12-04T08:53:09.5412437Z  * [new branch]                gh/jansel/559/orig      -> origin/gh/jansel/559/orig
2025-12-04T08:53:09.5412504Z  * [new branch]                gh/jansel/560/base      -> origin/gh/jansel/560/base
2025-12-04T08:53:09.5412574Z  * [new branch]                gh/jansel/560/head      -> origin/gh/jansel/560/head
2025-12-04T08:53:09.5412641Z  * [new branch]                gh/jansel/560/orig      -> origin/gh/jansel/560/orig
2025-12-04T08:53:09.5412707Z  * [new branch]                gh/jansel/561/base      -> origin/gh/jansel/561/base
2025-12-04T08:53:09.5412776Z  * [new branch]                gh/jansel/561/head      -> origin/gh/jansel/561/head
2025-12-04T08:53:09.5412844Z  * [new branch]                gh/jansel/561/orig      -> origin/gh/jansel/561/orig
2025-12-04T08:53:09.5412911Z  * [new branch]                gh/jansel/562/base      -> origin/gh/jansel/562/base
2025-12-04T08:53:09.5412981Z  * [new branch]                gh/jansel/562/head      -> origin/gh/jansel/562/head
2025-12-04T08:53:09.5413047Z  * [new branch]                gh/jansel/562/orig      -> origin/gh/jansel/562/orig
2025-12-04T08:53:09.5413114Z  * [new branch]                gh/jansel/563/base      -> origin/gh/jansel/563/base
2025-12-04T08:53:09.5413182Z  * [new branch]                gh/jansel/563/head      -> origin/gh/jansel/563/head
2025-12-04T08:53:09.5413250Z  * [new branch]                gh/jansel/563/orig      -> origin/gh/jansel/563/orig
2025-12-04T08:53:09.5413316Z  * [new branch]                gh/jansel/564/base      -> origin/gh/jansel/564/base
2025-12-04T08:53:09.5413385Z  * [new branch]                gh/jansel/564/head      -> origin/gh/jansel/564/head
2025-12-04T08:53:09.5413454Z  * [new branch]                gh/jansel/564/orig      -> origin/gh/jansel/564/orig
2025-12-04T08:53:09.5413522Z  * [new branch]                gh/jansel/565/base      -> origin/gh/jansel/565/base
2025-12-04T08:53:09.5413590Z  * [new branch]                gh/jansel/565/head      -> origin/gh/jansel/565/head
2025-12-04T08:53:09.5413658Z  * [new branch]                gh/jansel/565/orig      -> origin/gh/jansel/565/orig
2025-12-04T08:53:09.5413726Z  * [new branch]                gh/jansel/566/base      -> origin/gh/jansel/566/base
2025-12-04T08:53:09.5413793Z  * [new branch]                gh/jansel/566/head      -> origin/gh/jansel/566/head
2025-12-04T08:53:09.5413860Z  * [new branch]                gh/jansel/566/orig      -> origin/gh/jansel/566/orig
2025-12-04T08:53:09.5413928Z  * [new branch]                gh/jansel/567/base      -> origin/gh/jansel/567/base
2025-12-04T08:53:09.5413995Z  * [new branch]                gh/jansel/567/head      -> origin/gh/jansel/567/head
2025-12-04T08:53:09.5414063Z  * [new branch]                gh/jansel/567/orig      -> origin/gh/jansel/567/orig
2025-12-04T08:53:09.5414132Z  * [new branch]                gh/jansel/568/base      -> origin/gh/jansel/568/base
2025-12-04T08:53:09.5414200Z  * [new branch]                gh/jansel/568/head      -> origin/gh/jansel/568/head
2025-12-04T08:53:09.5414266Z  * [new branch]                gh/jansel/568/orig      -> origin/gh/jansel/568/orig
2025-12-04T08:53:09.5414335Z  * [new branch]                gh/jansel/569/base      -> origin/gh/jansel/569/base
2025-12-04T08:53:09.5414401Z  * [new branch]                gh/jansel/569/head      -> origin/gh/jansel/569/head
2025-12-04T08:53:09.5414468Z  * [new branch]                gh/jansel/569/orig      -> origin/gh/jansel/569/orig
2025-12-04T08:53:09.5414537Z  * [new branch]                gh/jansel/570/base      -> origin/gh/jansel/570/base
2025-12-04T08:53:09.5414605Z  * [new branch]                gh/jansel/570/head      -> origin/gh/jansel/570/head
2025-12-04T08:53:09.5414700Z  * [new branch]                gh/jansel/570/orig      -> origin/gh/jansel/570/orig
2025-12-04T08:53:09.5414768Z  * [new branch]                gh/jansel/571/base      -> origin/gh/jansel/571/base
2025-12-04T08:53:09.5414875Z  * [new branch]                gh/jansel/571/head      -> origin/gh/jansel/571/head
2025-12-04T08:53:09.5414942Z  * [new branch]                gh/jansel/571/orig      -> origin/gh/jansel/571/orig
2025-12-04T08:53:09.5415011Z  * [new branch]                gh/jansel/572/base      -> origin/gh/jansel/572/base
2025-12-04T08:53:09.5415078Z  * [new branch]                gh/jansel/572/head      -> origin/gh/jansel/572/head
2025-12-04T08:53:09.5415147Z  * [new branch]                gh/jansel/572/orig      -> origin/gh/jansel/572/orig
2025-12-04T08:53:09.5415214Z  * [new branch]                gh/jansel/573/base      -> origin/gh/jansel/573/base
2025-12-04T08:53:09.5415281Z  * [new branch]                gh/jansel/573/head      -> origin/gh/jansel/573/head
2025-12-04T08:53:09.5415353Z  * [new branch]                gh/jansel/573/orig      -> origin/gh/jansel/573/orig
2025-12-04T08:53:09.5415420Z  * [new branch]                gh/jansel/574/base      -> origin/gh/jansel/574/base
2025-12-04T08:53:09.5415488Z  * [new branch]                gh/jansel/574/head      -> origin/gh/jansel/574/head
2025-12-04T08:53:09.5415556Z  * [new branch]                gh/jansel/574/orig      -> origin/gh/jansel/574/orig
2025-12-04T08:53:09.5415623Z  * [new branch]                gh/jansel/575/base      -> origin/gh/jansel/575/base
2025-12-04T08:53:09.5415689Z  * [new branch]                gh/jansel/575/head      -> origin/gh/jansel/575/head
2025-12-04T08:53:09.5415760Z  * [new branch]                gh/jansel/575/orig      -> origin/gh/jansel/575/orig
2025-12-04T08:53:09.5415828Z  * [new branch]                gh/jansel/576/base      -> origin/gh/jansel/576/base
2025-12-04T08:53:09.5415895Z  * [new branch]                gh/jansel/576/head      -> origin/gh/jansel/576/head
2025-12-04T08:53:09.5415965Z  * [new branch]                gh/jansel/576/orig      -> origin/gh/jansel/576/orig
2025-12-04T08:53:09.5416046Z  * [new branch]                gh/jbschlosser/247/base -> origin/gh/jbschlosser/247/base
2025-12-04T08:53:09.5416128Z  * [new branch]                gh/jbschlosser/247/head -> origin/gh/jbschlosser/247/head
2025-12-04T08:53:09.5416206Z  * [new branch]                gh/jbschlosser/247/orig -> origin/gh/jbschlosser/247/orig
2025-12-04T08:53:09.5416283Z  * [new branch]                gh/jbschlosser/250/base -> origin/gh/jbschlosser/250/base
2025-12-04T08:53:09.5416359Z  * [new branch]                gh/jbschlosser/250/head -> origin/gh/jbschlosser/250/head
2025-12-04T08:53:09.5416435Z  * [new branch]                gh/jbschlosser/250/orig -> origin/gh/jbschlosser/250/orig
2025-12-04T08:53:09.5416508Z  * [new branch]                gh/jerryzh168/1/base    -> origin/gh/jerryzh168/1/base
2025-12-04T08:53:09.5416580Z  * [new branch]                gh/jerryzh168/1/head    -> origin/gh/jerryzh168/1/head
2025-12-04T08:53:09.5416654Z  * [new branch]                gh/jerryzh168/1/orig    -> origin/gh/jerryzh168/1/orig
2025-12-04T08:53:09.5416727Z  * [new branch]                gh/jiayisunx/59/base    -> origin/gh/jiayisunx/59/base
2025-12-04T08:53:09.5416801Z  * [new branch]                gh/jiayisunx/59/head    -> origin/gh/jiayisunx/59/head
2025-12-04T08:53:09.5416873Z  * [new branch]                gh/jiayisunx/59/orig    -> origin/gh/jiayisunx/59/orig
2025-12-04T08:53:09.5416944Z  * [new branch]                gh/jiayisunx/61/base    -> origin/gh/jiayisunx/61/base
2025-12-04T08:53:09.5417016Z  * [new branch]                gh/jiayisunx/61/head    -> origin/gh/jiayisunx/61/head
2025-12-04T08:53:09.5417090Z  * [new branch]                gh/jiayisunx/61/orig    -> origin/gh/jiayisunx/61/orig
2025-12-04T08:53:09.5417160Z  * [new branch]                gh/jiayisunx/68/base    -> origin/gh/jiayisunx/68/base
2025-12-04T08:53:09.5417234Z  * [new branch]                gh/jiayisunx/68/head    -> origin/gh/jiayisunx/68/head
2025-12-04T08:53:09.5417332Z  * [new branch]                gh/jiayisunx/68/orig    -> origin/gh/jiayisunx/68/orig
2025-12-04T08:53:09.5417402Z  * [new branch]                gh/jiayisunx/77/base    -> origin/gh/jiayisunx/77/base
2025-12-04T08:53:09.5417501Z  * [new branch]                gh/jiayisunx/77/head    -> origin/gh/jiayisunx/77/head
2025-12-04T08:53:09.5417572Z  * [new branch]                gh/jiayisunx/77/orig    -> origin/gh/jiayisunx/77/orig
2025-12-04T08:53:09.5417642Z  * [new branch]                gh/jiayisunx/78/base    -> origin/gh/jiayisunx/78/base
2025-12-04T08:53:09.5417713Z  * [new branch]                gh/jiayisunx/78/head    -> origin/gh/jiayisunx/78/head
2025-12-04T08:53:09.5417784Z  * [new branch]                gh/jiayisunx/78/orig    -> origin/gh/jiayisunx/78/orig
2025-12-04T08:53:09.5417854Z  * [new branch]                gh/jiayisunx/79/base    -> origin/gh/jiayisunx/79/base
2025-12-04T08:53:09.5417927Z  * [new branch]                gh/jiayisunx/79/head    -> origin/gh/jiayisunx/79/head
2025-12-04T08:53:09.5417999Z  * [new branch]                gh/jiayisunx/79/orig    -> origin/gh/jiayisunx/79/orig
2025-12-04T08:53:09.5418071Z  * [new branch]                gh/jiayisunx/82/base    -> origin/gh/jiayisunx/82/base
2025-12-04T08:53:09.5418145Z  * [new branch]                gh/jiayisunx/82/head    -> origin/gh/jiayisunx/82/head
2025-12-04T08:53:09.5418216Z  * [new branch]                gh/jiayisunx/82/orig    -> origin/gh/jiayisunx/82/orig
2025-12-04T08:53:09.5418288Z  * [new branch]                gh/jiayisunx/83/base    -> origin/gh/jiayisunx/83/base
2025-12-04T08:53:09.5418360Z  * [new branch]                gh/jiayisunx/83/head    -> origin/gh/jiayisunx/83/head
2025-12-04T08:53:09.5418431Z  * [new branch]                gh/jiayisunx/83/orig    -> origin/gh/jiayisunx/83/orig
2025-12-04T08:53:09.5418502Z  * [new branch]                gh/jiayisunx/84/base    -> origin/gh/jiayisunx/84/base
2025-12-04T08:53:09.5418572Z  * [new branch]                gh/jiayisunx/84/head    -> origin/gh/jiayisunx/84/head
2025-12-04T08:53:09.5418644Z  * [new branch]                gh/jiayisunx/84/orig    -> origin/gh/jiayisunx/84/orig
2025-12-04T08:53:09.5418717Z  * [new branch]                gh/jiayisunx/85/base    -> origin/gh/jiayisunx/85/base
2025-12-04T08:53:09.5418788Z  * [new branch]                gh/jiayisunx/85/head    -> origin/gh/jiayisunx/85/head
2025-12-04T08:53:09.5418858Z  * [new branch]                gh/jiayisunx/85/orig    -> origin/gh/jiayisunx/85/orig
2025-12-04T08:53:09.5418929Z  * [new branch]                gh/jiayisunx/86/base    -> origin/gh/jiayisunx/86/base
2025-12-04T08:53:09.5419000Z  * [new branch]                gh/jiayisunx/86/head    -> origin/gh/jiayisunx/86/head
2025-12-04T08:53:09.5419070Z  * [new branch]                gh/jiayisunx/86/orig    -> origin/gh/jiayisunx/86/orig
2025-12-04T08:53:09.5419144Z  * [new branch]                gh/jiayisunx/87/base    -> origin/gh/jiayisunx/87/base
2025-12-04T08:53:09.5419214Z  * [new branch]                gh/jiayisunx/87/head    -> origin/gh/jiayisunx/87/head
2025-12-04T08:53:09.5419286Z  * [new branch]                gh/jiayisunx/87/orig    -> origin/gh/jiayisunx/87/orig
2025-12-04T08:53:09.5419361Z  * [new branch]                gh/jiayisunx/88/base    -> origin/gh/jiayisunx/88/base
2025-12-04T08:53:09.5419431Z  * [new branch]                gh/jiayisunx/88/head    -> origin/gh/jiayisunx/88/head
2025-12-04T08:53:09.5419503Z  * [new branch]                gh/jiayisunx/88/orig    -> origin/gh/jiayisunx/88/orig
2025-12-04T08:53:09.5419575Z  * [new branch]                gh/jiayisunx/89/base    -> origin/gh/jiayisunx/89/base
2025-12-04T08:53:09.5419646Z  * [new branch]                gh/jiayisunx/89/head    -> origin/gh/jiayisunx/89/head
2025-12-04T08:53:09.5419717Z  * [new branch]                gh/jiayisunx/89/orig    -> origin/gh/jiayisunx/89/orig
2025-12-04T08:53:09.5419789Z  * [new branch]                gh/jiayisunx/90/base    -> origin/gh/jiayisunx/90/base
2025-12-04T08:53:09.5419861Z  * [new branch]                gh/jiayisunx/90/head    -> origin/gh/jiayisunx/90/head
2025-12-04T08:53:09.5419958Z  * [new branch]                gh/jiayisunx/90/orig    -> origin/gh/jiayisunx/90/orig
2025-12-04T08:53:09.5420070Z  * [new branch]                gh/jjwu@meta.com/1/base -> origin/gh/jjwu@meta.com/1/base
2025-12-04T08:53:09.5420147Z  * [new branch]                gh/jjwu@meta.com/1/head -> origin/gh/jjwu@meta.com/1/head
2025-12-04T08:53:09.5420219Z  * [new branch]                gh/jturney/1/base       -> origin/gh/jturney/1/base
2025-12-04T08:53:09.5420288Z  * [new branch]                gh/jturney/1/head       -> origin/gh/jturney/1/head
2025-12-04T08:53:09.5420357Z  * [new branch]                gh/jturney/1/orig       -> origin/gh/jturney/1/orig
2025-12-04T08:53:09.5420475Z  * [new branch]                gh/jturney/2/base       -> origin/gh/jturney/2/base
2025-12-04T08:53:09.5420541Z  * [new branch]                gh/jturney/2/head       -> origin/gh/jturney/2/head
2025-12-04T08:53:09.5420607Z  * [new branch]                gh/jturney/2/orig       -> origin/gh/jturney/2/orig
2025-12-04T08:53:09.5420691Z  * [new branch]                gh/karthickai/10/base   -> origin/gh/karthickai/10/base
2025-12-04T08:53:09.5420769Z  * [new branch]                gh/karthickai/10/head   -> origin/gh/karthickai/10/head
2025-12-04T08:53:09.5420844Z  * [new branch]                gh/karthickai/10/orig   -> origin/gh/karthickai/10/orig
2025-12-04T08:53:09.5420919Z  * [new branch]                gh/karthickai/11/base   -> origin/gh/karthickai/11/base
2025-12-04T08:53:09.5420993Z  * [new branch]                gh/karthickai/11/head   -> origin/gh/karthickai/11/head
2025-12-04T08:53:09.5421068Z  * [new branch]                gh/karthickai/11/orig   -> origin/gh/karthickai/11/orig
2025-12-04T08:53:09.5421144Z  * [new branch]                gh/karthickai/12/base   -> origin/gh/karthickai/12/base
2025-12-04T08:53:09.5421217Z  * [new branch]                gh/karthickai/12/head   -> origin/gh/karthickai/12/head
2025-12-04T08:53:09.5421291Z  * [new branch]                gh/karthickai/12/orig   -> origin/gh/karthickai/12/orig
2025-12-04T08:53:09.5421370Z  * [new branch]                gh/karthickai/13/base   -> origin/gh/karthickai/13/base
2025-12-04T08:53:09.5421448Z  * [new branch]                gh/karthickai/13/head   -> origin/gh/karthickai/13/head
2025-12-04T08:53:09.5421523Z  * [new branch]                gh/karthickai/13/orig   -> origin/gh/karthickai/13/orig
2025-12-04T08:53:09.5421596Z  * [new branch]                gh/karthickai/14/base   -> origin/gh/karthickai/14/base
2025-12-04T08:53:09.5421669Z  * [new branch]                gh/karthickai/14/head   -> origin/gh/karthickai/14/head
2025-12-04T08:53:09.5421743Z  * [new branch]                gh/karthickai/14/orig   -> origin/gh/karthickai/14/orig
2025-12-04T08:53:09.5421815Z  * [new branch]                gh/karthickai/15/base   -> origin/gh/karthickai/15/base
2025-12-04T08:53:09.5421888Z  * [new branch]                gh/karthickai/15/head   -> origin/gh/karthickai/15/head
2025-12-04T08:53:09.5421967Z  * [new branch]                gh/karthickai/15/orig   -> origin/gh/karthickai/15/orig
2025-12-04T08:53:09.5422039Z  * [new branch]                gh/karthickai/16/base   -> origin/gh/karthickai/16/base
2025-12-04T08:53:09.5422113Z  * [new branch]                gh/karthickai/16/head   -> origin/gh/karthickai/16/head
2025-12-04T08:53:09.5422190Z  * [new branch]                gh/karthickai/16/orig   -> origin/gh/karthickai/16/orig
2025-12-04T08:53:09.5422263Z  * [new branch]                gh/karthickai/17/base   -> origin/gh/karthickai/17/base
2025-12-04T08:53:09.5422337Z  * [new branch]                gh/karthickai/17/head   -> origin/gh/karthickai/17/head
2025-12-04T08:53:09.5422414Z  * [new branch]                gh/karthickai/17/orig   -> origin/gh/karthickai/17/orig
2025-12-04T08:53:09.5422487Z  * [new branch]                gh/karthickai/18/base   -> origin/gh/karthickai/18/base
2025-12-04T08:53:09.5422561Z  * [new branch]                gh/karthickai/18/head   -> origin/gh/karthickai/18/head
2025-12-04T08:53:09.5422678Z  * [new branch]                gh/karthickai/18/orig   -> origin/gh/karthickai/18/orig
2025-12-04T08:53:09.5422750Z  * [new branch]                gh/karthickai/19/base   -> origin/gh/karthickai/19/base
2025-12-04T08:53:09.5422862Z  * [new branch]                gh/karthickai/19/head   -> origin/gh/karthickai/19/head
2025-12-04T08:53:09.5422938Z  * [new branch]                gh/karthickai/19/orig   -> origin/gh/karthickai/19/orig
2025-12-04T08:53:09.5423012Z  * [new branch]                gh/karthickai/20/base   -> origin/gh/karthickai/20/base
2025-12-04T08:53:09.5423088Z  * [new branch]                gh/karthickai/20/head   -> origin/gh/karthickai/20/head
2025-12-04T08:53:09.5423160Z  * [new branch]                gh/karthickai/20/orig   -> origin/gh/karthickai/20/orig
2025-12-04T08:53:09.5423233Z  * [new branch]                gh/karthickai/21/base   -> origin/gh/karthickai/21/base
2025-12-04T08:53:09.5423308Z  * [new branch]                gh/karthickai/21/head   -> origin/gh/karthickai/21/head
2025-12-04T08:53:09.5423383Z  * [new branch]                gh/karthickai/21/orig   -> origin/gh/karthickai/21/orig
2025-12-04T08:53:09.5423458Z  * [new branch]                gh/karthickai/22/base   -> origin/gh/karthickai/22/base
2025-12-04T08:53:09.5423531Z  * [new branch]                gh/karthickai/22/head   -> origin/gh/karthickai/22/head
2025-12-04T08:53:09.5423605Z  * [new branch]                gh/karthickai/22/orig   -> origin/gh/karthickai/22/orig
2025-12-04T08:53:09.5423680Z  * [new branch]                gh/karthickai/23/base   -> origin/gh/karthickai/23/base
2025-12-04T08:53:09.5423755Z  * [new branch]                gh/karthickai/23/head   -> origin/gh/karthickai/23/head
2025-12-04T08:53:09.5423828Z  * [new branch]                gh/karthickai/23/orig   -> origin/gh/karthickai/23/orig
2025-12-04T08:53:09.5423902Z  * [new branch]                gh/karthickai/24/base   -> origin/gh/karthickai/24/base
2025-12-04T08:53:09.5423980Z  * [new branch]                gh/karthickai/24/head   -> origin/gh/karthickai/24/head
2025-12-04T08:53:09.5424055Z  * [new branch]                gh/karthickai/24/orig   -> origin/gh/karthickai/24/orig
2025-12-04T08:53:09.5424130Z  * [new branch]                gh/karthickai/25/base   -> origin/gh/karthickai/25/base
2025-12-04T08:53:09.5424206Z  * [new branch]                gh/karthickai/25/head   -> origin/gh/karthickai/25/head
2025-12-04T08:53:09.5424279Z  * [new branch]                gh/karthickai/25/orig   -> origin/gh/karthickai/25/orig
2025-12-04T08:53:09.5424351Z  * [new branch]                gh/karthickai/26/base   -> origin/gh/karthickai/26/base
2025-12-04T08:53:09.5424427Z  * [new branch]                gh/karthickai/26/head   -> origin/gh/karthickai/26/head
2025-12-04T08:53:09.5424502Z  * [new branch]                gh/karthickai/26/orig   -> origin/gh/karthickai/26/orig
2025-12-04T08:53:09.5424576Z  * [new branch]                gh/karthickai/6/base    -> origin/gh/karthickai/6/base
2025-12-04T08:53:09.5424653Z  * [new branch]                gh/karthickai/6/head    -> origin/gh/karthickai/6/head
2025-12-04T08:53:09.5424726Z  * [new branch]                gh/karthickai/6/orig    -> origin/gh/karthickai/6/orig
2025-12-04T08:53:09.5424800Z  * [new branch]                gh/krocki/1/base        -> origin/gh/krocki/1/base
2025-12-04T08:53:09.5424869Z  * [new branch]                gh/krocki/1/head        -> origin/gh/krocki/1/head
2025-12-04T08:53:09.5424936Z  * [new branch]                gh/krocki/1/orig        -> origin/gh/krocki/1/orig
2025-12-04T08:53:09.5425004Z  * [new branch]                gh/krocki/2/base        -> origin/gh/krocki/2/base
2025-12-04T08:53:09.5425071Z  * [new branch]                gh/krocki/2/head        -> origin/gh/krocki/2/head
2025-12-04T08:53:09.5425135Z  * [new branch]                gh/krocki/2/orig        -> origin/gh/krocki/2/orig
2025-12-04T08:53:09.5425217Z  * [new branch]                gh/kurtamohler/60/base  -> origin/gh/kurtamohler/60/base
2025-12-04T08:53:09.5425294Z  * [new branch]                gh/kurtamohler/60/head  -> origin/gh/kurtamohler/60/head
2025-12-04T08:53:09.5425396Z  * [new branch]                gh/kurtamohler/60/orig  -> origin/gh/kurtamohler/60/orig
2025-12-04T08:53:09.5425513Z  * [new branch]                gh/kurtamohler/61/base  -> origin/gh/kurtamohler/61/base
2025-12-04T08:53:09.5425589Z  * [new branch]                gh/kurtamohler/61/head  -> origin/gh/kurtamohler/61/head
2025-12-04T08:53:09.5425662Z  * [new branch]                gh/kurtamohler/61/orig  -> origin/gh/kurtamohler/61/orig
2025-12-04T08:53:09.5425742Z  * [new branch]                gh/kurtamohler/62/base  -> origin/gh/kurtamohler/62/base
2025-12-04T08:53:09.5425816Z  * [new branch]                gh/kurtamohler/62/head  -> origin/gh/kurtamohler/62/head
2025-12-04T08:53:09.5425891Z  * [new branch]                gh/kurtamohler/62/orig  -> origin/gh/kurtamohler/62/orig
2025-12-04T08:53:09.5425968Z  * [new branch]                gh/kurtamohler/63/base  -> origin/gh/kurtamohler/63/base
2025-12-04T08:53:09.5426043Z  * [new branch]                gh/kurtamohler/63/head  -> origin/gh/kurtamohler/63/head
2025-12-04T08:53:09.5426118Z  * [new branch]                gh/kurtamohler/63/orig  -> origin/gh/kurtamohler/63/orig
2025-12-04T08:53:09.5426197Z  * [new branch]                gh/kurtamohler/64/base  -> origin/gh/kurtamohler/64/base
2025-12-04T08:53:09.5426271Z  * [new branch]                gh/kurtamohler/64/head  -> origin/gh/kurtamohler/64/head
2025-12-04T08:53:09.5426348Z  * [new branch]                gh/kurtamohler/64/orig  -> origin/gh/kurtamohler/64/orig
2025-12-04T08:53:09.5426423Z  * [new branch]                gh/kurtamohler/65/base  -> origin/gh/kurtamohler/65/base
2025-12-04T08:53:09.5426498Z  * [new branch]                gh/kurtamohler/65/head  -> origin/gh/kurtamohler/65/head
2025-12-04T08:53:09.5426574Z  * [new branch]                gh/kurtamohler/65/orig  -> origin/gh/kurtamohler/65/orig
2025-12-04T08:53:09.5426649Z  * [new branch]                gh/kurtamohler/66/base  -> origin/gh/kurtamohler/66/base
2025-12-04T08:53:09.5426725Z  * [new branch]                gh/kurtamohler/66/head  -> origin/gh/kurtamohler/66/head
2025-12-04T08:53:09.5426803Z  * [new branch]                gh/kurtamohler/66/orig  -> origin/gh/kurtamohler/66/orig
2025-12-04T08:53:09.5426880Z  * [new branch]                gh/kurtamohler/67/base  -> origin/gh/kurtamohler/67/base
2025-12-04T08:53:09.5426955Z  * [new branch]                gh/kurtamohler/67/head  -> origin/gh/kurtamohler/67/head
2025-12-04T08:53:09.5427034Z  * [new branch]                gh/kurtamohler/67/orig  -> origin/gh/kurtamohler/67/orig
2025-12-04T08:53:09.5427105Z  * [new branch]                gh/kwen2501/130/base    -> origin/gh/kwen2501/130/base
2025-12-04T08:53:09.5427175Z  * [new branch]                gh/kwen2501/130/head    -> origin/gh/kwen2501/130/head
2025-12-04T08:53:09.5427247Z  * [new branch]                gh/kwen2501/130/orig    -> origin/gh/kwen2501/130/orig
2025-12-04T08:53:09.5427316Z  * [new branch]                gh/kwen2501/170/base    -> origin/gh/kwen2501/170/base
2025-12-04T08:53:09.5427390Z  * [new branch]                gh/kwen2501/170/head    -> origin/gh/kwen2501/170/head
2025-12-04T08:53:09.5427462Z  * [new branch]                gh/kwen2501/187/base    -> origin/gh/kwen2501/187/base
2025-12-04T08:53:09.5427534Z  * [new branch]                gh/kwen2501/187/head    -> origin/gh/kwen2501/187/head
2025-12-04T08:53:09.5427607Z  * [new branch]                gh/kwen2501/187/orig    -> origin/gh/kwen2501/187/orig
2025-12-04T08:53:09.5427675Z  * [new branch]                gh/kwen2501/188/base    -> origin/gh/kwen2501/188/base
2025-12-04T08:53:09.5427745Z  * [new branch]                gh/kwen2501/188/head    -> origin/gh/kwen2501/188/head
2025-12-04T08:53:09.5427814Z  * [new branch]                gh/kwen2501/188/orig    -> origin/gh/kwen2501/188/orig
2025-12-04T08:53:09.5427883Z  * [new branch]                gh/kwen2501/211/base    -> origin/gh/kwen2501/211/base
2025-12-04T08:53:09.5427954Z  * [new branch]                gh/kwen2501/211/head    -> origin/gh/kwen2501/211/head
2025-12-04T08:53:09.5428055Z  * [new branch]                gh/kwen2501/224/base    -> origin/gh/kwen2501/224/base
2025-12-04T08:53:09.5428151Z  * [new branch]                gh/kwen2501/224/head    -> origin/gh/kwen2501/224/head
2025-12-04T08:53:09.5428223Z  * [new branch]                gh/kwen2501/224/orig    -> origin/gh/kwen2501/224/orig
2025-12-04T08:53:09.5428292Z  * [new branch]                gh/kwen2501/228/base    -> origin/gh/kwen2501/228/base
2025-12-04T08:53:09.5428360Z  * [new branch]                gh/kwen2501/228/head    -> origin/gh/kwen2501/228/head
2025-12-04T08:53:09.5428430Z  * [new branch]                gh/kwen2501/228/orig    -> origin/gh/kwen2501/228/orig
2025-12-04T08:53:09.5428500Z  * [new branch]                gh/kwen2501/234/base    -> origin/gh/kwen2501/234/base
2025-12-04T08:53:09.5428569Z  * [new branch]                gh/kwen2501/234/head    -> origin/gh/kwen2501/234/head
2025-12-04T08:53:09.5428639Z  * [new branch]                gh/kwen2501/234/orig    -> origin/gh/kwen2501/234/orig
2025-12-04T08:53:09.5428709Z  * [new branch]                gh/kwen2501/235/base    -> origin/gh/kwen2501/235/base
2025-12-04T08:53:09.5428782Z  * [new branch]                gh/kwen2501/235/head    -> origin/gh/kwen2501/235/head
2025-12-04T08:53:09.5428852Z  * [new branch]                gh/kwen2501/235/orig    -> origin/gh/kwen2501/235/orig
2025-12-04T08:53:09.5428921Z  * [new branch]                gh/kwen2501/236/base    -> origin/gh/kwen2501/236/base
2025-12-04T08:53:09.5428991Z  * [new branch]                gh/kwen2501/236/head    -> origin/gh/kwen2501/236/head
2025-12-04T08:53:09.5429060Z  * [new branch]                gh/kwen2501/236/orig    -> origin/gh/kwen2501/236/orig
2025-12-04T08:53:09.5429129Z  * [new branch]                gh/kwen2501/237/base    -> origin/gh/kwen2501/237/base
2025-12-04T08:53:09.5429198Z  * [new branch]                gh/kwen2501/237/head    -> origin/gh/kwen2501/237/head
2025-12-04T08:53:09.5429267Z  * [new branch]                gh/kwen2501/237/orig    -> origin/gh/kwen2501/237/orig
2025-12-04T08:53:09.5429338Z  * [new branch]                gh/kwen2501/238/base    -> origin/gh/kwen2501/238/base
2025-12-04T08:53:09.5429413Z  * [new branch]                gh/kwen2501/238/head    -> origin/gh/kwen2501/238/head
2025-12-04T08:53:09.5429483Z  * [new branch]                gh/kwen2501/238/orig    -> origin/gh/kwen2501/238/orig
2025-12-04T08:53:09.5429551Z  * [new branch]                gh/kwen2501/240/base    -> origin/gh/kwen2501/240/base
2025-12-04T08:53:09.5429622Z  * [new branch]                gh/kwen2501/240/head    -> origin/gh/kwen2501/240/head
2025-12-04T08:53:09.5429690Z  * [new branch]                gh/kwen2501/240/orig    -> origin/gh/kwen2501/240/orig
2025-12-04T08:53:09.5429759Z  * [new branch]                gh/kwen2501/241/base    -> origin/gh/kwen2501/241/base
2025-12-04T08:53:09.5429830Z  * [new branch]                gh/kwen2501/241/head    -> origin/gh/kwen2501/241/head
2025-12-04T08:53:09.5429900Z  * [new branch]                gh/kwen2501/241/orig    -> origin/gh/kwen2501/241/orig
2025-12-04T08:53:09.5429971Z  * [new branch]                gh/kwen2501/247/base    -> origin/gh/kwen2501/247/base
2025-12-04T08:53:09.5430044Z  * [new branch]                gh/kwen2501/247/head    -> origin/gh/kwen2501/247/head
2025-12-04T08:53:09.5430114Z  * [new branch]                gh/kwen2501/247/orig    -> origin/gh/kwen2501/247/orig
2025-12-04T08:53:09.5430184Z  * [new branch]                gh/kwen2501/252/base    -> origin/gh/kwen2501/252/base
2025-12-04T08:53:09.5430253Z  * [new branch]                gh/kwen2501/252/head    -> origin/gh/kwen2501/252/head
2025-12-04T08:53:09.5430321Z  * [new branch]                gh/kwen2501/252/orig    -> origin/gh/kwen2501/252/orig
2025-12-04T08:53:09.5430392Z  * [new branch]                gh/kwen2501/259/base    -> origin/gh/kwen2501/259/base
2025-12-04T08:53:09.5430551Z  * [new branch]                gh/kwen2501/259/head    -> origin/gh/kwen2501/259/head
2025-12-04T08:53:09.5430620Z  * [new branch]                gh/kwen2501/259/orig    -> origin/gh/kwen2501/259/orig
2025-12-04T08:53:09.5430733Z  * [new branch]                gh/kwen2501/260/base    -> origin/gh/kwen2501/260/base
2025-12-04T08:53:09.5430856Z  * [new branch]                gh/kwen2501/260/head    -> origin/gh/kwen2501/260/head
2025-12-04T08:53:09.5430926Z  * [new branch]                gh/kwen2501/260/orig    -> origin/gh/kwen2501/260/orig
2025-12-04T08:53:09.5430995Z  * [new branch]                gh/kwen2501/268/base    -> origin/gh/kwen2501/268/base
2025-12-04T08:53:09.5431064Z  * [new branch]                gh/kwen2501/268/head    -> origin/gh/kwen2501/268/head
2025-12-04T08:53:09.5431132Z  * [new branch]                gh/kwen2501/268/orig    -> origin/gh/kwen2501/268/orig
2025-12-04T08:53:09.5431201Z  * [new branch]                gh/kwen2501/269/base    -> origin/gh/kwen2501/269/base
2025-12-04T08:53:09.5431270Z  * [new branch]                gh/kwen2501/269/head    -> origin/gh/kwen2501/269/head
2025-12-04T08:53:09.5431338Z  * [new branch]                gh/kwen2501/269/orig    -> origin/gh/kwen2501/269/orig
2025-12-04T08:53:09.5431410Z  * [new branch]                gh/kwen2501/270/base    -> origin/gh/kwen2501/270/base
2025-12-04T08:53:09.5431481Z  * [new branch]                gh/kwen2501/270/head    -> origin/gh/kwen2501/270/head
2025-12-04T08:53:09.5431550Z  * [new branch]                gh/kwen2501/270/orig    -> origin/gh/kwen2501/270/orig
2025-12-04T08:53:09.5431619Z  * [new branch]                gh/kwen2501/271/base    -> origin/gh/kwen2501/271/base
2025-12-04T08:53:09.5431687Z  * [new branch]                gh/kwen2501/271/head    -> origin/gh/kwen2501/271/head
2025-12-04T08:53:09.5431757Z  * [new branch]                gh/kwen2501/271/orig    -> origin/gh/kwen2501/271/orig
2025-12-04T08:53:09.5431826Z  * [new branch]                gh/kwen2501/274/base    -> origin/gh/kwen2501/274/base
2025-12-04T08:53:09.5431894Z  * [new branch]                gh/kwen2501/274/head    -> origin/gh/kwen2501/274/head
2025-12-04T08:53:09.5431967Z  * [new branch]                gh/kwen2501/274/orig    -> origin/gh/kwen2501/274/orig
2025-12-04T08:53:09.5432036Z  * [new branch]                gh/kwen2501/275/base    -> origin/gh/kwen2501/275/base
2025-12-04T08:53:09.5432108Z  * [new branch]                gh/kwen2501/275/head    -> origin/gh/kwen2501/275/head
2025-12-04T08:53:09.5432178Z  * [new branch]                gh/kwen2501/275/orig    -> origin/gh/kwen2501/275/orig
2025-12-04T08:53:09.5432247Z  * [new branch]                gh/kwen2501/276/base    -> origin/gh/kwen2501/276/base
2025-12-04T08:53:09.5432316Z  * [new branch]                gh/kwen2501/276/head    -> origin/gh/kwen2501/276/head
2025-12-04T08:53:09.5432385Z  * [new branch]                gh/kwen2501/276/orig    -> origin/gh/kwen2501/276/orig
2025-12-04T08:53:09.5432454Z  * [new branch]                gh/kwen2501/277/base    -> origin/gh/kwen2501/277/base
2025-12-04T08:53:09.5432523Z  * [new branch]                gh/kwen2501/277/head    -> origin/gh/kwen2501/277/head
2025-12-04T08:53:09.5432598Z  * [new branch]                gh/kwen2501/277/orig    -> origin/gh/kwen2501/277/orig
2025-12-04T08:53:09.5432666Z  * [new branch]                gh/kwen2501/278/base    -> origin/gh/kwen2501/278/base
2025-12-04T08:53:09.5432736Z  * [new branch]                gh/kwen2501/278/head    -> origin/gh/kwen2501/278/head
2025-12-04T08:53:09.5432806Z  * [new branch]                gh/kwen2501/278/orig    -> origin/gh/kwen2501/278/orig
2025-12-04T08:53:09.5432876Z  * [new branch]                gh/kwen2501/279/base    -> origin/gh/kwen2501/279/base
2025-12-04T08:53:09.5432944Z  * [new branch]                gh/kwen2501/279/head    -> origin/gh/kwen2501/279/head
2025-12-04T08:53:09.5433014Z  * [new branch]                gh/kwen2501/279/orig    -> origin/gh/kwen2501/279/orig
2025-12-04T08:53:09.5433082Z  * [new branch]                gh/kwen2501/280/base    -> origin/gh/kwen2501/280/base
2025-12-04T08:53:09.5433151Z  * [new branch]                gh/kwen2501/280/head    -> origin/gh/kwen2501/280/head
2025-12-04T08:53:09.5433250Z  * [new branch]                gh/kwen2501/280/orig    -> origin/gh/kwen2501/280/orig
2025-12-04T08:53:09.5433318Z  * [new branch]                gh/kwen2501/281/base    -> origin/gh/kwen2501/281/base
2025-12-04T08:53:09.5433413Z  * [new branch]                gh/kwen2501/281/head    -> origin/gh/kwen2501/281/head
2025-12-04T08:53:09.5433482Z  * [new branch]                gh/kwen2501/281/orig    -> origin/gh/kwen2501/281/orig
2025-12-04T08:53:09.5433550Z  * [new branch]                gh/kwen2501/282/base    -> origin/gh/kwen2501/282/base
2025-12-04T08:53:09.5433621Z  * [new branch]                gh/kwen2501/282/head    -> origin/gh/kwen2501/282/head
2025-12-04T08:53:09.5433690Z  * [new branch]                gh/kwen2501/282/orig    -> origin/gh/kwen2501/282/orig
2025-12-04T08:53:09.5433759Z  * [new branch]                gh/kwen2501/283/base    -> origin/gh/kwen2501/283/base
2025-12-04T08:53:09.5433829Z  * [new branch]                gh/kwen2501/283/head    -> origin/gh/kwen2501/283/head
2025-12-04T08:53:09.5433898Z  * [new branch]                gh/kwen2501/283/orig    -> origin/gh/kwen2501/283/orig
2025-12-04T08:53:09.5433967Z  * [new branch]                gh/kwen2501/284/base    -> origin/gh/kwen2501/284/base
2025-12-04T08:53:09.5434038Z  * [new branch]                gh/kwen2501/284/head    -> origin/gh/kwen2501/284/head
2025-12-04T08:53:09.5434106Z  * [new branch]                gh/kwen2501/284/orig    -> origin/gh/kwen2501/284/orig
2025-12-04T08:53:09.5434175Z  * [new branch]                gh/kwen2501/285/base    -> origin/gh/kwen2501/285/base
2025-12-04T08:53:09.5434245Z  * [new branch]                gh/kwen2501/285/head    -> origin/gh/kwen2501/285/head
2025-12-04T08:53:09.5434314Z  * [new branch]                gh/kwen2501/285/orig    -> origin/gh/kwen2501/285/orig
2025-12-04T08:53:09.5434383Z  * [new branch]                gh/kwen2501/286/base    -> origin/gh/kwen2501/286/base
2025-12-04T08:53:09.5434452Z  * [new branch]                gh/kwen2501/286/head    -> origin/gh/kwen2501/286/head
2025-12-04T08:53:09.5434522Z  * [new branch]                gh/kwen2501/286/orig    -> origin/gh/kwen2501/286/orig
2025-12-04T08:53:09.5434591Z  * [new branch]                gh/kwen2501/287/base    -> origin/gh/kwen2501/287/base
2025-12-04T08:53:09.5434662Z  * [new branch]                gh/kwen2501/287/head    -> origin/gh/kwen2501/287/head
2025-12-04T08:53:09.5434731Z  * [new branch]                gh/kwen2501/287/orig    -> origin/gh/kwen2501/287/orig
2025-12-04T08:53:09.5434801Z  * [new branch]                gh/kwen2501/288/base    -> origin/gh/kwen2501/288/base
2025-12-04T08:53:09.5434869Z  * [new branch]                gh/kwen2501/288/head    -> origin/gh/kwen2501/288/head
2025-12-04T08:53:09.5434937Z  * [new branch]                gh/kwen2501/288/orig    -> origin/gh/kwen2501/288/orig
2025-12-04T08:53:09.5435014Z  * [new branch]                gh/laithsakka/251/base  -> origin/gh/laithsakka/251/base
2025-12-04T08:53:09.5435089Z  * [new branch]                gh/laithsakka/251/head  -> origin/gh/laithsakka/251/head
2025-12-04T08:53:09.5435163Z  * [new branch]                gh/laithsakka/251/orig  -> origin/gh/laithsakka/251/orig
2025-12-04T08:53:09.5435240Z  * [new branch]                gh/laithsakka/276/base  -> origin/gh/laithsakka/276/base
2025-12-04T08:53:09.5435313Z  * [new branch]                gh/laithsakka/276/head  -> origin/gh/laithsakka/276/head
2025-12-04T08:53:09.5435386Z  * [new branch]                gh/laithsakka/276/orig  -> origin/gh/laithsakka/276/orig
2025-12-04T08:53:09.5435461Z  * [new branch]                gh/laithsakka/28/base   -> origin/gh/laithsakka/28/base
2025-12-04T08:53:09.5435534Z  * [new branch]                gh/laithsakka/29/base   -> origin/gh/laithsakka/29/base
2025-12-04T08:53:09.5435607Z  * [new branch]                gh/laithsakka/30/base   -> origin/gh/laithsakka/30/base
2025-12-04T08:53:09.5435681Z  * [new branch]                gh/laithsakka/30/head   -> origin/gh/laithsakka/30/head
2025-12-04T08:53:09.5435755Z  * [new branch]                gh/laithsakka/31/base   -> origin/gh/laithsakka/31/base
2025-12-04T08:53:09.5435861Z  * [new branch]                gh/laithsakka/31/head   -> origin/gh/laithsakka/31/head
2025-12-04T08:53:09.5435958Z  * [new branch]                gh/laithsakka/313/base  -> origin/gh/laithsakka/313/base
2025-12-04T08:53:09.5436032Z  * [new branch]                gh/laithsakka/313/head  -> origin/gh/laithsakka/313/head
2025-12-04T08:53:09.5436105Z  * [new branch]                gh/laithsakka/313/orig  -> origin/gh/laithsakka/313/orig
2025-12-04T08:53:09.5436180Z  * [new branch]                gh/laithsakka/316/base  -> origin/gh/laithsakka/316/base
2025-12-04T08:53:09.5436254Z  * [new branch]                gh/laithsakka/316/head  -> origin/gh/laithsakka/316/head
2025-12-04T08:53:09.5436328Z  * [new branch]                gh/laithsakka/316/orig  -> origin/gh/laithsakka/316/orig
2025-12-04T08:53:09.5436400Z  * [new branch]                gh/laithsakka/317/base  -> origin/gh/laithsakka/317/base
2025-12-04T08:53:09.5436474Z  * [new branch]                gh/laithsakka/317/head  -> origin/gh/laithsakka/317/head
2025-12-04T08:53:09.5436549Z  * [new branch]                gh/laithsakka/317/orig  -> origin/gh/laithsakka/317/orig
2025-12-04T08:53:09.5436623Z  * [new branch]                gh/laithsakka/319/base  -> origin/gh/laithsakka/319/base
2025-12-04T08:53:09.5436696Z  * [new branch]                gh/laithsakka/319/head  -> origin/gh/laithsakka/319/head
2025-12-04T08:53:09.5436770Z  * [new branch]                gh/laithsakka/319/orig  -> origin/gh/laithsakka/319/orig
2025-12-04T08:53:09.5436843Z  * [new branch]                gh/laithsakka/32/base   -> origin/gh/laithsakka/32/base
2025-12-04T08:53:09.5436916Z  * [new branch]                gh/laithsakka/32/head   -> origin/gh/laithsakka/32/head
2025-12-04T08:53:09.5436991Z  * [new branch]                gh/laithsakka/320/base  -> origin/gh/laithsakka/320/base
2025-12-04T08:53:09.5437064Z  * [new branch]                gh/laithsakka/320/head  -> origin/gh/laithsakka/320/head
2025-12-04T08:53:09.5437139Z  * [new branch]                gh/laithsakka/320/orig  -> origin/gh/laithsakka/320/orig
2025-12-04T08:53:09.5437213Z  * [new branch]                gh/laithsakka/321/base  -> origin/gh/laithsakka/321/base
2025-12-04T08:53:09.5437288Z  * [new branch]                gh/laithsakka/321/head  -> origin/gh/laithsakka/321/head
2025-12-04T08:53:09.5437361Z  * [new branch]                gh/laithsakka/321/orig  -> origin/gh/laithsakka/321/orig
2025-12-04T08:53:09.5437436Z  * [new branch]                gh/laithsakka/322/base  -> origin/gh/laithsakka/322/base
2025-12-04T08:53:09.5437510Z  * [new branch]                gh/laithsakka/322/head  -> origin/gh/laithsakka/322/head
2025-12-04T08:53:09.5437582Z  * [new branch]                gh/laithsakka/322/orig  -> origin/gh/laithsakka/322/orig
2025-12-04T08:53:09.5437656Z  * [new branch]                gh/laithsakka/323/base  -> origin/gh/laithsakka/323/base
2025-12-04T08:53:09.5437729Z  * [new branch]                gh/laithsakka/323/head  -> origin/gh/laithsakka/323/head
2025-12-04T08:53:09.5437803Z  * [new branch]                gh/laithsakka/323/orig  -> origin/gh/laithsakka/323/orig
2025-12-04T08:53:09.5437878Z  * [new branch]                gh/laithsakka/324/base  -> origin/gh/laithsakka/324/base
2025-12-04T08:53:09.5437951Z  * [new branch]                gh/laithsakka/324/head  -> origin/gh/laithsakka/324/head
2025-12-04T08:53:09.5438025Z  * [new branch]                gh/laithsakka/324/orig  -> origin/gh/laithsakka/324/orig
2025-12-04T08:53:09.5438098Z  * [new branch]                gh/laithsakka/325/base  -> origin/gh/laithsakka/325/base
2025-12-04T08:53:09.5438172Z  * [new branch]                gh/laithsakka/325/head  -> origin/gh/laithsakka/325/head
2025-12-04T08:53:09.5438246Z  * [new branch]                gh/laithsakka/325/orig  -> origin/gh/laithsakka/325/orig
2025-12-04T08:53:09.5438319Z  * [new branch]                gh/laithsakka/326/base  -> origin/gh/laithsakka/326/base
2025-12-04T08:53:09.5438391Z  * [new branch]                gh/laithsakka/326/head  -> origin/gh/laithsakka/326/head
2025-12-04T08:53:09.5438490Z  * [new branch]                gh/laithsakka/326/orig  -> origin/gh/laithsakka/326/orig
2025-12-04T08:53:09.5438584Z  * [new branch]                gh/laithsakka/327/base  -> origin/gh/laithsakka/327/base
2025-12-04T08:53:09.5438658Z  * [new branch]                gh/laithsakka/327/head  -> origin/gh/laithsakka/327/head
2025-12-04T08:53:09.5438732Z  * [new branch]                gh/laithsakka/327/orig  -> origin/gh/laithsakka/327/orig
2025-12-04T08:53:09.5438805Z  * [new branch]                gh/laithsakka/328/base  -> origin/gh/laithsakka/328/base
2025-12-04T08:53:09.5438878Z  * [new branch]                gh/laithsakka/328/head  -> origin/gh/laithsakka/328/head
2025-12-04T08:53:09.5438952Z  * [new branch]                gh/laithsakka/328/orig  -> origin/gh/laithsakka/328/orig
2025-12-04T08:53:09.5439022Z  * [new branch]                gh/liangel/4/base       -> origin/gh/liangel/4/base
2025-12-04T08:53:09.5439091Z  * [new branch]                gh/liangel/4/head       -> origin/gh/liangel/4/head
2025-12-04T08:53:09.5439162Z  * [new branch]                gh/liangel/4/orig       -> origin/gh/liangel/4/orig
2025-12-04T08:53:09.5439240Z  * [new branch]                gh/lucaskabela/1/base   -> origin/gh/lucaskabela/1/base
2025-12-04T08:53:09.5439314Z  * [new branch]                gh/lucaskabela/1/head   -> origin/gh/lucaskabela/1/head
2025-12-04T08:53:09.5439381Z  * [new branch]                gh/lw/4/base            -> origin/gh/lw/4/base
2025-12-04T08:53:09.5439443Z  * [new branch]                gh/lw/4/head            -> origin/gh/lw/4/head
2025-12-04T08:53:09.5439507Z  * [new branch]                gh/lw/4/orig            -> origin/gh/lw/4/orig
2025-12-04T08:53:09.5439567Z  * [new branch]                gh/lw/5/base            -> origin/gh/lw/5/base
2025-12-04T08:53:09.5439628Z  * [new branch]                gh/lw/5/head            -> origin/gh/lw/5/head
2025-12-04T08:53:09.5439689Z  * [new branch]                gh/lw/5/orig            -> origin/gh/lw/5/orig
2025-12-04T08:53:09.5439750Z  * [new branch]                gh/lw/6/base            -> origin/gh/lw/6/base
2025-12-04T08:53:09.5439813Z  * [new branch]                gh/lw/6/head            -> origin/gh/lw/6/head
2025-12-04T08:53:09.5439876Z  * [new branch]                gh/lw/6/orig            -> origin/gh/lw/6/orig
2025-12-04T08:53:09.5439944Z  * [new branch]                gh/malfet/14/base       -> origin/gh/malfet/14/base
2025-12-04T08:53:09.5440015Z  * [new branch]                gh/malfet/417/base      -> origin/gh/malfet/417/base
2025-12-04T08:53:09.5440085Z  * [new branch]                gh/malfet/417/head      -> origin/gh/malfet/417/head
2025-12-04T08:53:09.5440154Z  * [new branch]                gh/malfet/417/orig      -> origin/gh/malfet/417/orig
2025-12-04T08:53:09.5440221Z  * [new branch]                gh/malfet/506/base      -> origin/gh/malfet/506/base
2025-12-04T08:53:09.5440290Z  * [new branch]                gh/malfet/506/head      -> origin/gh/malfet/506/head
2025-12-04T08:53:09.5440360Z  * [new branch]                gh/malfet/506/orig      -> origin/gh/malfet/506/orig
2025-12-04T08:53:09.5440454Z  * [new branch]                gh/malfet/517/base      -> origin/gh/malfet/517/base
2025-12-04T08:53:09.5440524Z  * [new branch]                gh/malfet/517/head      -> origin/gh/malfet/517/head
2025-12-04T08:53:09.5440591Z  * [new branch]                gh/malfet/528/base      -> origin/gh/malfet/528/base
2025-12-04T08:53:09.5440657Z  * [new branch]                gh/malfet/528/head      -> origin/gh/malfet/528/head
2025-12-04T08:53:09.5440725Z  * [new branch]                gh/malfet/528/orig      -> origin/gh/malfet/528/orig
2025-12-04T08:53:09.5440792Z  * [new branch]                gh/malfet/537/base      -> origin/gh/malfet/537/base
2025-12-04T08:53:09.5440858Z  * [new branch]                gh/malfet/537/head      -> origin/gh/malfet/537/head
2025-12-04T08:53:09.5440927Z  * [new branch]                gh/malfet/537/orig      -> origin/gh/malfet/537/orig
2025-12-04T08:53:09.5441038Z  * [new branch]                gh/malfet/546/base      -> origin/gh/malfet/546/base
2025-12-04T08:53:09.5441106Z  * [new branch]                gh/malfet/546/head      -> origin/gh/malfet/546/head
2025-12-04T08:53:09.5441220Z  * [new branch]                gh/malfet/546/orig      -> origin/gh/malfet/546/orig
2025-12-04T08:53:09.5441288Z  * [new branch]                gh/malfet/565/base      -> origin/gh/malfet/565/base
2025-12-04T08:53:09.5441357Z  * [new branch]                gh/malfet/565/head      -> origin/gh/malfet/565/head
2025-12-04T08:53:09.5441424Z  * [new branch]                gh/malfet/565/orig      -> origin/gh/malfet/565/orig
2025-12-04T08:53:09.5441491Z  * [new branch]                gh/malfet/575/base      -> origin/gh/malfet/575/base
2025-12-04T08:53:09.5441559Z  * [new branch]                gh/malfet/575/head      -> origin/gh/malfet/575/head
2025-12-04T08:53:09.5441626Z  * [new branch]                gh/malfet/575/orig      -> origin/gh/malfet/575/orig
2025-12-04T08:53:09.5441693Z  * [new branch]                gh/malfet/580/base      -> origin/gh/malfet/580/base
2025-12-04T08:53:09.5441761Z  * [new branch]                gh/malfet/580/head      -> origin/gh/malfet/580/head
2025-12-04T08:53:09.5441829Z  * [new branch]                gh/malfet/580/orig      -> origin/gh/malfet/580/orig
2025-12-04T08:53:09.5441896Z  * [new branch]                gh/malfet/581/base      -> origin/gh/malfet/581/base
2025-12-04T08:53:09.5441964Z  * [new branch]                gh/malfet/581/head      -> origin/gh/malfet/581/head
2025-12-04T08:53:09.5442031Z  * [new branch]                gh/malfet/581/orig      -> origin/gh/malfet/581/orig
2025-12-04T08:53:09.5442097Z  * [new branch]                gh/malfet/583/base      -> origin/gh/malfet/583/base
2025-12-04T08:53:09.5442165Z  * [new branch]                gh/malfet/583/head      -> origin/gh/malfet/583/head
2025-12-04T08:53:09.5442231Z  * [new branch]                gh/malfet/583/orig      -> origin/gh/malfet/583/orig
2025-12-04T08:53:09.5442298Z  * [new branch]                gh/malfet/586/base      -> origin/gh/malfet/586/base
2025-12-04T08:53:09.5442366Z  * [new branch]                gh/malfet/586/head      -> origin/gh/malfet/586/head
2025-12-04T08:53:09.5442434Z  * [new branch]                gh/malfet/586/orig      -> origin/gh/malfet/586/orig
2025-12-04T08:53:09.5442500Z  * [new branch]                gh/malfet/587/base      -> origin/gh/malfet/587/base
2025-12-04T08:53:09.5442567Z  * [new branch]                gh/malfet/587/head      -> origin/gh/malfet/587/head
2025-12-04T08:53:09.5442634Z  * [new branch]                gh/malfet/587/orig      -> origin/gh/malfet/587/orig
2025-12-04T08:53:09.5442700Z  * [new branch]                gh/malfet/588/base      -> origin/gh/malfet/588/base
2025-12-04T08:53:09.5442769Z  * [new branch]                gh/malfet/588/head      -> origin/gh/malfet/588/head
2025-12-04T08:53:09.5442836Z  * [new branch]                gh/malfet/588/orig      -> origin/gh/malfet/588/orig
2025-12-04T08:53:09.5442905Z  * [new branch]                gh/malfet/589/base      -> origin/gh/malfet/589/base
2025-12-04T08:53:09.5442972Z  * [new branch]                gh/malfet/589/head      -> origin/gh/malfet/589/head
2025-12-04T08:53:09.5443040Z  * [new branch]                gh/malfet/589/orig      -> origin/gh/malfet/589/orig
2025-12-04T08:53:09.5443108Z  * [new branch]                gh/malfet/590/base      -> origin/gh/malfet/590/base
2025-12-04T08:53:09.5443175Z  * [new branch]                gh/malfet/590/head      -> origin/gh/malfet/590/head
2025-12-04T08:53:09.5443241Z  * [new branch]                gh/malfet/590/orig      -> origin/gh/malfet/590/orig
2025-12-04T08:53:09.5443308Z  * [new branch]                gh/malfet/591/base      -> origin/gh/malfet/591/base
2025-12-04T08:53:09.5443375Z  * [new branch]                gh/malfet/591/head      -> origin/gh/malfet/591/head
2025-12-04T08:53:09.5443441Z  * [new branch]                gh/malfet/591/orig      -> origin/gh/malfet/591/orig
2025-12-04T08:53:09.5443532Z  * [new branch]                gh/malfet/592/base      -> origin/gh/malfet/592/base
2025-12-04T08:53:09.5443599Z  * [new branch]                gh/malfet/592/head      -> origin/gh/malfet/592/head
2025-12-04T08:53:09.5443703Z  * [new branch]                gh/malfet/592/orig      -> origin/gh/malfet/592/orig
2025-12-04T08:53:09.5443772Z  * [new branch]                gh/malfet/593/base      -> origin/gh/malfet/593/base
2025-12-04T08:53:09.5443838Z  * [new branch]                gh/malfet/593/head      -> origin/gh/malfet/593/head
2025-12-04T08:53:09.5443905Z  * [new branch]                gh/malfet/593/orig      -> origin/gh/malfet/593/orig
2025-12-04T08:53:09.5443973Z  * [new branch]                gh/malfet/594/base      -> origin/gh/malfet/594/base
2025-12-04T08:53:09.5444039Z  * [new branch]                gh/malfet/594/head      -> origin/gh/malfet/594/head
2025-12-04T08:53:09.5444106Z  * [new branch]                gh/malfet/594/orig      -> origin/gh/malfet/594/orig
2025-12-04T08:53:09.5444174Z  * [new branch]                gh/malfet/595/base      -> origin/gh/malfet/595/base
2025-12-04T08:53:09.5444242Z  * [new branch]                gh/malfet/595/head      -> origin/gh/malfet/595/head
2025-12-04T08:53:09.5444310Z  * [new branch]                gh/malfet/595/orig      -> origin/gh/malfet/595/orig
2025-12-04T08:53:09.5444378Z  * [new branch]                gh/malfet/596/base      -> origin/gh/malfet/596/base
2025-12-04T08:53:09.5444445Z  * [new branch]                gh/malfet/596/head      -> origin/gh/malfet/596/head
2025-12-04T08:53:09.5444513Z  * [new branch]                gh/malfet/596/orig      -> origin/gh/malfet/596/orig
2025-12-04T08:53:09.5444579Z  * [new branch]                gh/malfet/597/base      -> origin/gh/malfet/597/base
2025-12-04T08:53:09.5444645Z  * [new branch]                gh/malfet/597/head      -> origin/gh/malfet/597/head
2025-12-04T08:53:09.5444715Z  * [new branch]                gh/malfet/597/orig      -> origin/gh/malfet/597/orig
2025-12-04T08:53:09.5444782Z  * [new branch]                gh/malfet/598/base      -> origin/gh/malfet/598/base
2025-12-04T08:53:09.5444851Z  * [new branch]                gh/malfet/598/head      -> origin/gh/malfet/598/head
2025-12-04T08:53:09.5444921Z  * [new branch]                gh/malfet/598/orig      -> origin/gh/malfet/598/orig
2025-12-04T08:53:09.5444987Z  * [new branch]                gh/malfet/599/base      -> origin/gh/malfet/599/base
2025-12-04T08:53:09.5445054Z  * [new branch]                gh/malfet/599/head      -> origin/gh/malfet/599/head
2025-12-04T08:53:09.5445124Z  * [new branch]                gh/malfet/599/orig      -> origin/gh/malfet/599/orig
2025-12-04T08:53:09.5445191Z  * [new branch]                gh/malfet/600/base      -> origin/gh/malfet/600/base
2025-12-04T08:53:09.5445258Z  * [new branch]                gh/malfet/600/head      -> origin/gh/malfet/600/head
2025-12-04T08:53:09.5445327Z  * [new branch]                gh/malfet/600/orig      -> origin/gh/malfet/600/orig
2025-12-04T08:53:09.5445394Z  * [new branch]                gh/malfet/601/base      -> origin/gh/malfet/601/base
2025-12-04T08:53:09.5445462Z  * [new branch]                gh/malfet/601/head      -> origin/gh/malfet/601/head
2025-12-04T08:53:09.5445531Z  * [new branch]                gh/malfet/601/orig      -> origin/gh/malfet/601/orig
2025-12-04T08:53:09.5445598Z  * [new branch]                gh/malfet/602/base      -> origin/gh/malfet/602/base
2025-12-04T08:53:09.5445666Z  * [new branch]                gh/malfet/602/head      -> origin/gh/malfet/602/head
2025-12-04T08:53:09.5445734Z  * [new branch]                gh/malfet/602/orig      -> origin/gh/malfet/602/orig
2025-12-04T08:53:09.5445802Z  * [new branch]                gh/malfet/603/base      -> origin/gh/malfet/603/base
2025-12-04T08:53:09.5445869Z  * [new branch]                gh/malfet/603/head      -> origin/gh/malfet/603/head
2025-12-04T08:53:09.5445937Z  * [new branch]                gh/malfet/603/orig      -> origin/gh/malfet/603/orig
2025-12-04T08:53:09.5446004Z  * [new branch]                gh/malfet/604/base      -> origin/gh/malfet/604/base
2025-12-04T08:53:09.5446655Z  * [new branch]                gh/malfet/604/head      -> origin/gh/malfet/604/head
2025-12-04T08:53:09.5446745Z  * [new branch]                gh/malfet/604/orig      -> origin/gh/malfet/604/orig
2025-12-04T08:53:09.5446812Z  * [new branch]                gh/malfet/605/base      -> origin/gh/malfet/605/base
2025-12-04T08:53:09.5446880Z  * [new branch]                gh/malfet/605/head      -> origin/gh/malfet/605/head
2025-12-04T08:53:09.5446948Z  * [new branch]                gh/malfet/605/orig      -> origin/gh/malfet/605/orig
2025-12-04T08:53:09.5447015Z  * [new branch]                gh/malfet/606/base      -> origin/gh/malfet/606/base
2025-12-04T08:53:09.5447082Z  * [new branch]                gh/malfet/606/head      -> origin/gh/malfet/606/head
2025-12-04T08:53:09.5447149Z  * [new branch]                gh/malfet/606/orig      -> origin/gh/malfet/606/orig
2025-12-04T08:53:09.5447217Z  * [new branch]                gh/malfet/607/base      -> origin/gh/malfet/607/base
2025-12-04T08:53:09.5447291Z  * [new branch]                gh/malfet/607/head      -> origin/gh/malfet/607/head
2025-12-04T08:53:09.5447358Z  * [new branch]                gh/malfet/607/orig      -> origin/gh/malfet/607/orig
2025-12-04T08:53:09.5447426Z  * [new branch]                gh/malfet/608/base      -> origin/gh/malfet/608/base
2025-12-04T08:53:09.5447494Z  * [new branch]                gh/malfet/608/head      -> origin/gh/malfet/608/head
2025-12-04T08:53:09.5447561Z  * [new branch]                gh/malfet/608/orig      -> origin/gh/malfet/608/orig
2025-12-04T08:53:09.5447628Z  * [new branch]                gh/malfet/609/base      -> origin/gh/malfet/609/base
2025-12-04T08:53:09.5447696Z  * [new branch]                gh/malfet/609/head      -> origin/gh/malfet/609/head
2025-12-04T08:53:09.5447763Z  * [new branch]                gh/malfet/609/orig      -> origin/gh/malfet/609/orig
2025-12-04T08:53:09.5447831Z  * [new branch]                gh/malfet/610/base      -> origin/gh/malfet/610/base
2025-12-04T08:53:09.5447901Z  * [new branch]                gh/malfet/610/head      -> origin/gh/malfet/610/head
2025-12-04T08:53:09.5447969Z  * [new branch]                gh/malfet/610/orig      -> origin/gh/malfet/610/orig
2025-12-04T08:53:09.5448036Z  * [new branch]                gh/malfet/611/base      -> origin/gh/malfet/611/base
2025-12-04T08:53:09.5448104Z  * [new branch]                gh/malfet/611/head      -> origin/gh/malfet/611/head
2025-12-04T08:53:09.5448172Z  * [new branch]                gh/malfet/611/orig      -> origin/gh/malfet/611/orig
2025-12-04T08:53:09.5448238Z  * [new branch]                gh/malfet/612/base      -> origin/gh/malfet/612/base
2025-12-04T08:53:09.5448306Z  * [new branch]                gh/malfet/612/head      -> origin/gh/malfet/612/head
2025-12-04T08:53:09.5448373Z  * [new branch]                gh/malfet/612/orig      -> origin/gh/malfet/612/orig
2025-12-04T08:53:09.5448442Z  * [new branch]                gh/malfet/64/base       -> origin/gh/malfet/64/base
2025-12-04T08:53:09.5448510Z  * [new branch]                gh/malfet/64/head       -> origin/gh/malfet/64/head
2025-12-04T08:53:09.5448602Z  * [new branch]                gh/manuelcandales/11/base -> origin/gh/manuelcandales/11/base
2025-12-04T08:53:09.5448689Z  * [new branch]                gh/manuelcandales/11/head -> origin/gh/manuelcandales/11/head
2025-12-04T08:53:09.5448772Z  * [new branch]                gh/manuelcandales/11/orig -> origin/gh/manuelcandales/11/orig
2025-12-04T08:53:09.5448840Z  * [new branch]                gh/markkm/1/base        -> origin/gh/markkm/1/base
2025-12-04T08:53:09.5448913Z  * [new branch]                gh/masnesral/1/base     -> origin/gh/masnesral/1/base
2025-12-04T08:53:09.5448985Z  * [new branch]                gh/masnesral/1/head     -> origin/gh/masnesral/1/head
2025-12-04T08:53:09.5449057Z  * [new branch]                gh/masnesral/1/orig     -> origin/gh/masnesral/1/orig
2025-12-04T08:53:09.5449128Z  * [new branch]                gh/mhorowitz/0/base     -> origin/gh/mhorowitz/0/base
2025-12-04T08:53:09.5449226Z  * [new branch]                gh/mhorowitz/0/head     -> origin/gh/mhorowitz/0/head
2025-12-04T08:53:09.5449324Z  * [new branch]                gh/mhorowitz/1/base     -> origin/gh/mhorowitz/1/base
2025-12-04T08:53:09.5449395Z  * [new branch]                gh/mhorowitz/1/head     -> origin/gh/mhorowitz/1/head
2025-12-04T08:53:09.5449464Z  * [new branch]                gh/mhorowitz/2/base     -> origin/gh/mhorowitz/2/base
2025-12-04T08:53:09.5449533Z  * [new branch]                gh/mhorowitz/2/head     -> origin/gh/mhorowitz/2/head
2025-12-04T08:53:09.5449604Z  * [new branch]                gh/mhorowitz/3/base     -> origin/gh/mhorowitz/3/base
2025-12-04T08:53:09.5449673Z  * [new branch]                gh/mhorowitz/3/head     -> origin/gh/mhorowitz/3/head
2025-12-04T08:53:09.5449742Z  * [new branch]                gh/mhorowitz/4/base     -> origin/gh/mhorowitz/4/base
2025-12-04T08:53:09.5449812Z  * [new branch]                gh/mhorowitz/4/head     -> origin/gh/mhorowitz/4/head
2025-12-04T08:53:09.5449883Z  * [new branch]                gh/mhorowitz/5/base     -> origin/gh/mhorowitz/5/base
2025-12-04T08:53:09.5449953Z  * [new branch]                gh/mhorowitz/5/head     -> origin/gh/mhorowitz/5/head
2025-12-04T08:53:09.5450024Z  * [new branch]                gh/mhorowitz/6/base     -> origin/gh/mhorowitz/6/base
2025-12-04T08:53:09.5450092Z  * [new branch]                gh/mhorowitz/6/head     -> origin/gh/mhorowitz/6/head
2025-12-04T08:53:09.5450194Z  * [new branch]                gh/mikaylagawarecki/234/base -> origin/gh/mikaylagawarecki/234/base
2025-12-04T08:53:09.5450292Z  * [new branch]                gh/mikaylagawarecki/234/head -> origin/gh/mikaylagawarecki/234/head
2025-12-04T08:53:09.5450386Z  * [new branch]                gh/mikaylagawarecki/235/base -> origin/gh/mikaylagawarecki/235/base
2025-12-04T08:53:09.5450511Z  * [new branch]                gh/mikaylagawarecki/235/head -> origin/gh/mikaylagawarecki/235/head
2025-12-04T08:53:09.5450604Z  * [new branch]                gh/mikaylagawarecki/236/base -> origin/gh/mikaylagawarecki/236/base
2025-12-04T08:53:09.5450696Z  * [new branch]                gh/mikaylagawarecki/236/head -> origin/gh/mikaylagawarecki/236/head
2025-12-04T08:53:09.5450788Z  * [new branch]                gh/mikaylagawarecki/237/base -> origin/gh/mikaylagawarecki/237/base
2025-12-04T08:53:09.5450880Z  * [new branch]                gh/mikaylagawarecki/237/head -> origin/gh/mikaylagawarecki/237/head
2025-12-04T08:53:09.5450971Z  * [new branch]                gh/mikaylagawarecki/238/base -> origin/gh/mikaylagawarecki/238/base
2025-12-04T08:53:09.5451063Z  * [new branch]                gh/mikaylagawarecki/238/head -> origin/gh/mikaylagawarecki/238/head
2025-12-04T08:53:09.5451155Z  * [new branch]                gh/mikaylagawarecki/336/base -> origin/gh/mikaylagawarecki/336/base
2025-12-04T08:53:09.5451245Z  * [new branch]                gh/mikaylagawarecki/336/head -> origin/gh/mikaylagawarecki/336/head
2025-12-04T08:53:09.5451341Z  * [new branch]                gh/mikaylagawarecki/336/orig -> origin/gh/mikaylagawarecki/336/orig
2025-12-04T08:53:09.5451434Z  * [new branch]                gh/mikaylagawarecki/341/base -> origin/gh/mikaylagawarecki/341/base
2025-12-04T08:53:09.5451526Z  * [new branch]                gh/mikaylagawarecki/341/head -> origin/gh/mikaylagawarecki/341/head
2025-12-04T08:53:09.5451617Z  * [new branch]                gh/mikaylagawarecki/341/orig -> origin/gh/mikaylagawarecki/341/orig
2025-12-04T08:53:09.5451707Z  * [new branch]                gh/mikaylagawarecki/342/base -> origin/gh/mikaylagawarecki/342/base
2025-12-04T08:53:09.5451799Z  * [new branch]                gh/mikaylagawarecki/342/head -> origin/gh/mikaylagawarecki/342/head
2025-12-04T08:53:09.5451890Z  * [new branch]                gh/mikaylagawarecki/342/orig -> origin/gh/mikaylagawarecki/342/orig
2025-12-04T08:53:09.5451980Z  * [new branch]                gh/mikaylagawarecki/345/base -> origin/gh/mikaylagawarecki/345/base
2025-12-04T08:53:09.5452116Z  * [new branch]                gh/mikaylagawarecki/345/head -> origin/gh/mikaylagawarecki/345/head
2025-12-04T08:53:09.5452249Z  * [new branch]                gh/mikaylagawarecki/345/orig -> origin/gh/mikaylagawarecki/345/orig
2025-12-04T08:53:09.5452339Z  * [new branch]                gh/mikaylagawarecki/346/base -> origin/gh/mikaylagawarecki/346/base
2025-12-04T08:53:09.5452431Z  * [new branch]                gh/mikaylagawarecki/346/head -> origin/gh/mikaylagawarecki/346/head
2025-12-04T08:53:09.5452521Z  * [new branch]                gh/mikaylagawarecki/346/orig -> origin/gh/mikaylagawarecki/346/orig
2025-12-04T08:53:09.5452611Z  * [new branch]                gh/mikaylagawarecki/347/base -> origin/gh/mikaylagawarecki/347/base
2025-12-04T08:53:09.5452702Z  * [new branch]                gh/mikaylagawarecki/347/head -> origin/gh/mikaylagawarecki/347/head
2025-12-04T08:53:09.5452793Z  * [new branch]                gh/mikaylagawarecki/347/orig -> origin/gh/mikaylagawarecki/347/orig
2025-12-04T08:53:09.5452887Z  * [new branch]                gh/mikaylagawarecki/350/base -> origin/gh/mikaylagawarecki/350/base
2025-12-04T08:53:09.5452981Z  * [new branch]                gh/mikaylagawarecki/350/head -> origin/gh/mikaylagawarecki/350/head
2025-12-04T08:53:09.5453072Z  * [new branch]                gh/mikaylagawarecki/350/orig -> origin/gh/mikaylagawarecki/350/orig
2025-12-04T08:53:09.5453164Z  * [new branch]                gh/mikaylagawarecki/351/base -> origin/gh/mikaylagawarecki/351/base
2025-12-04T08:53:09.5453254Z  * [new branch]                gh/mikaylagawarecki/351/head -> origin/gh/mikaylagawarecki/351/head
2025-12-04T08:53:09.5453344Z  * [new branch]                gh/mikaylagawarecki/351/orig -> origin/gh/mikaylagawarecki/351/orig
2025-12-04T08:53:09.5453435Z  * [new branch]                gh/mikaylagawarecki/352/base -> origin/gh/mikaylagawarecki/352/base
2025-12-04T08:53:09.5453525Z  * [new branch]                gh/mikaylagawarecki/352/head -> origin/gh/mikaylagawarecki/352/head
2025-12-04T08:53:09.5453620Z  * [new branch]                gh/mikaylagawarecki/352/orig -> origin/gh/mikaylagawarecki/352/orig
2025-12-04T08:53:09.5453716Z  * [new branch]                gh/mikaylagawarecki/353/base -> origin/gh/mikaylagawarecki/353/base
2025-12-04T08:53:09.5453809Z  * [new branch]                gh/mikaylagawarecki/353/head -> origin/gh/mikaylagawarecki/353/head
2025-12-04T08:53:09.5453900Z  * [new branch]                gh/mikaylagawarecki/353/orig -> origin/gh/mikaylagawarecki/353/orig
2025-12-04T08:53:09.5453992Z  * [new branch]                gh/mikaylagawarecki/354/base -> origin/gh/mikaylagawarecki/354/base
2025-12-04T08:53:09.5454083Z  * [new branch]                gh/mikaylagawarecki/354/head -> origin/gh/mikaylagawarecki/354/head
2025-12-04T08:53:09.5454174Z  * [new branch]                gh/mikaylagawarecki/354/orig -> origin/gh/mikaylagawarecki/354/orig
2025-12-04T08:53:09.5454264Z  * [new branch]                gh/mikaylagawarecki/356/base -> origin/gh/mikaylagawarecki/356/base
2025-12-04T08:53:09.5454356Z  * [new branch]                gh/mikaylagawarecki/356/head -> origin/gh/mikaylagawarecki/356/head
2025-12-04T08:53:09.5454451Z  * [new branch]                gh/mikaylagawarecki/356/orig -> origin/gh/mikaylagawarecki/356/orig
2025-12-04T08:53:09.5454541Z  * [new branch]                gh/mikaylagawarecki/357/base -> origin/gh/mikaylagawarecki/357/base
2025-12-04T08:53:09.5454631Z  * [new branch]                gh/mikaylagawarecki/357/head -> origin/gh/mikaylagawarecki/357/head
2025-12-04T08:53:09.5454722Z  * [new branch]                gh/mikaylagawarecki/357/orig -> origin/gh/mikaylagawarecki/357/orig
2025-12-04T08:53:09.5454812Z  * [new branch]                gh/mikaylagawarecki/359/base -> origin/gh/mikaylagawarecki/359/base
2025-12-04T08:53:09.5454904Z  * [new branch]                gh/mikaylagawarecki/359/head -> origin/gh/mikaylagawarecki/359/head
2025-12-04T08:53:09.5454997Z  * [new branch]                gh/mikaylagawarecki/359/orig -> origin/gh/mikaylagawarecki/359/orig
2025-12-04T08:53:09.5455113Z  * [new branch]                gh/mikaylagawarecki/360/base -> origin/gh/mikaylagawarecki/360/base
2025-12-04T08:53:09.5455237Z  * [new branch]                gh/mikaylagawarecki/360/head -> origin/gh/mikaylagawarecki/360/head
2025-12-04T08:53:09.5455329Z  * [new branch]                gh/mikaylagawarecki/360/orig -> origin/gh/mikaylagawarecki/360/orig
2025-12-04T08:53:09.5455422Z  * [new branch]                gh/mikaylagawarecki/361/base -> origin/gh/mikaylagawarecki/361/base
2025-12-04T08:53:09.5455512Z  * [new branch]                gh/mikaylagawarecki/361/head -> origin/gh/mikaylagawarecki/361/head
2025-12-04T08:53:09.5455604Z  * [new branch]                gh/mikaylagawarecki/361/orig -> origin/gh/mikaylagawarecki/361/orig
2025-12-04T08:53:09.5455695Z  * [new branch]                gh/mikaylagawarecki/362/base -> origin/gh/mikaylagawarecki/362/base
2025-12-04T08:53:09.5455785Z  * [new branch]                gh/mikaylagawarecki/362/head -> origin/gh/mikaylagawarecki/362/head
2025-12-04T08:53:09.5455880Z  * [new branch]                gh/mikaylagawarecki/362/orig -> origin/gh/mikaylagawarecki/362/orig
2025-12-04T08:53:09.5455970Z  * [new branch]                gh/mikaylagawarecki/363/base -> origin/gh/mikaylagawarecki/363/base
2025-12-04T08:53:09.5456062Z  * [new branch]                gh/mikaylagawarecki/363/head -> origin/gh/mikaylagawarecki/363/head
2025-12-04T08:53:09.5456152Z  * [new branch]                gh/mikaylagawarecki/363/orig -> origin/gh/mikaylagawarecki/363/orig
2025-12-04T08:53:09.5456242Z  * [new branch]                gh/mikaylagawarecki/364/base -> origin/gh/mikaylagawarecki/364/base
2025-12-04T08:53:09.5456333Z  * [new branch]                gh/mikaylagawarecki/364/head -> origin/gh/mikaylagawarecki/364/head
2025-12-04T08:53:09.5456423Z  * [new branch]                gh/mikaylagawarecki/364/orig -> origin/gh/mikaylagawarecki/364/orig
2025-12-04T08:53:09.5456513Z  * [new branch]                gh/mikaylagawarecki/365/base -> origin/gh/mikaylagawarecki/365/base
2025-12-04T08:53:09.5456607Z  * [new branch]                gh/mikaylagawarecki/365/head -> origin/gh/mikaylagawarecki/365/head
2025-12-04T08:53:09.5456698Z  * [new branch]                gh/mikaylagawarecki/365/orig -> origin/gh/mikaylagawarecki/365/orig
2025-12-04T08:53:09.5456789Z  * [new branch]                gh/mikaylagawarecki/366/base -> origin/gh/mikaylagawarecki/366/base
2025-12-04T08:53:09.5456881Z  * [new branch]                gh/mikaylagawarecki/366/head -> origin/gh/mikaylagawarecki/366/head
2025-12-04T08:53:09.5456974Z  * [new branch]                gh/mikaylagawarecki/366/orig -> origin/gh/mikaylagawarecki/366/orig
2025-12-04T08:53:09.5457066Z  * [new branch]                gh/mikaylagawarecki/367/base -> origin/gh/mikaylagawarecki/367/base
2025-12-04T08:53:09.5457161Z  * [new branch]                gh/mikaylagawarecki/367/head -> origin/gh/mikaylagawarecki/367/head
2025-12-04T08:53:09.5457253Z  * [new branch]                gh/mikaylagawarecki/367/orig -> origin/gh/mikaylagawarecki/367/orig
2025-12-04T08:53:09.5457348Z  * [new branch]                gh/mikaylagawarecki/368/base -> origin/gh/mikaylagawarecki/368/base
2025-12-04T08:53:09.5457442Z  * [new branch]                gh/mikaylagawarecki/368/head -> origin/gh/mikaylagawarecki/368/head
2025-12-04T08:53:09.5457532Z  * [new branch]                gh/mikaylagawarecki/368/orig -> origin/gh/mikaylagawarecki/368/orig
2025-12-04T08:53:09.5457625Z  * [new branch]                gh/mikaylagawarecki/369/base -> origin/gh/mikaylagawarecki/369/base
2025-12-04T08:53:09.5457717Z  * [new branch]                gh/mikaylagawarecki/369/head -> origin/gh/mikaylagawarecki/369/head
2025-12-04T08:53:09.5457809Z  * [new branch]                gh/mikaylagawarecki/369/orig -> origin/gh/mikaylagawarecki/369/orig
2025-12-04T08:53:09.5457903Z  * [new branch]                gh/mikaylagawarecki/370/base -> origin/gh/mikaylagawarecki/370/base
2025-12-04T08:53:09.5457995Z  * [new branch]                gh/mikaylagawarecki/370/head -> origin/gh/mikaylagawarecki/370/head
2025-12-04T08:53:09.5458112Z  * [new branch]                gh/mikaylagawarecki/370/orig -> origin/gh/mikaylagawarecki/370/orig
2025-12-04T08:53:09.5458229Z  * [new branch]                gh/mikaylagawarecki/371/base -> origin/gh/mikaylagawarecki/371/base
2025-12-04T08:53:09.5458320Z  * [new branch]                gh/mikaylagawarecki/371/head -> origin/gh/mikaylagawarecki/371/head
2025-12-04T08:53:09.5458412Z  * [new branch]                gh/mikaylagawarecki/371/orig -> origin/gh/mikaylagawarecki/371/orig
2025-12-04T08:53:09.5458503Z  * [new branch]                gh/mikaylagawarecki/372/base -> origin/gh/mikaylagawarecki/372/base
2025-12-04T08:53:09.5458595Z  * [new branch]                gh/mikaylagawarecki/372/head -> origin/gh/mikaylagawarecki/372/head
2025-12-04T08:53:09.5458687Z  * [new branch]                gh/mikaylagawarecki/372/orig -> origin/gh/mikaylagawarecki/372/orig
2025-12-04T08:53:09.5458783Z  * [new branch]                gh/mikaylagawarecki/373/base -> origin/gh/mikaylagawarecki/373/base
2025-12-04T08:53:09.5458875Z  * [new branch]                gh/mikaylagawarecki/373/head -> origin/gh/mikaylagawarecki/373/head
2025-12-04T08:53:09.5458969Z  * [new branch]                gh/mikaylagawarecki/373/orig -> origin/gh/mikaylagawarecki/373/orig
2025-12-04T08:53:09.5459059Z  * [new branch]                gh/mikaylagawarecki/374/base -> origin/gh/mikaylagawarecki/374/base
2025-12-04T08:53:09.5459149Z  * [new branch]                gh/mikaylagawarecki/374/head -> origin/gh/mikaylagawarecki/374/head
2025-12-04T08:53:09.5459242Z  * [new branch]                gh/mikaylagawarecki/374/orig -> origin/gh/mikaylagawarecki/374/orig
2025-12-04T08:53:09.5459334Z  * [new branch]                gh/mikaylagawarecki/375/base -> origin/gh/mikaylagawarecki/375/base
2025-12-04T08:53:09.5459427Z  * [new branch]                gh/mikaylagawarecki/375/head -> origin/gh/mikaylagawarecki/375/head
2025-12-04T08:53:09.5459522Z  * [new branch]                gh/mikaylagawarecki/375/orig -> origin/gh/mikaylagawarecki/375/orig
2025-12-04T08:53:09.5459617Z  * [new branch]                gh/mikaylagawarecki/376/base -> origin/gh/mikaylagawarecki/376/base
2025-12-04T08:53:09.5459710Z  * [new branch]                gh/mikaylagawarecki/376/head -> origin/gh/mikaylagawarecki/376/head
2025-12-04T08:53:09.5459805Z  * [new branch]                gh/mikaylagawarecki/376/orig -> origin/gh/mikaylagawarecki/376/orig
2025-12-04T08:53:09.5459897Z  * [new branch]                gh/mikaylagawarecki/377/base -> origin/gh/mikaylagawarecki/377/base
2025-12-04T08:53:09.5459988Z  * [new branch]                gh/mikaylagawarecki/377/head -> origin/gh/mikaylagawarecki/377/head
2025-12-04T08:53:09.5460083Z  * [new branch]                gh/mikaylagawarecki/377/orig -> origin/gh/mikaylagawarecki/377/orig
2025-12-04T08:53:09.5460175Z  * [new branch]                gh/mikaylagawarecki/378/base -> origin/gh/mikaylagawarecki/378/base
2025-12-04T08:53:09.5460269Z  * [new branch]                gh/mikaylagawarecki/378/head -> origin/gh/mikaylagawarecki/378/head
2025-12-04T08:53:09.5460362Z  * [new branch]                gh/mikaylagawarecki/378/orig -> origin/gh/mikaylagawarecki/378/orig
2025-12-04T08:53:09.5460488Z  * [new branch]                gh/mikaylagawarecki/379/base -> origin/gh/mikaylagawarecki/379/base
2025-12-04T08:53:09.5460584Z  * [new branch]                gh/mikaylagawarecki/379/head -> origin/gh/mikaylagawarecki/379/head
2025-12-04T08:53:09.5460675Z  * [new branch]                gh/mikaylagawarecki/379/orig -> origin/gh/mikaylagawarecki/379/orig
2025-12-04T08:53:09.5460766Z  * [new branch]                gh/mikaylagawarecki/380/base -> origin/gh/mikaylagawarecki/380/base
2025-12-04T08:53:09.5460859Z  * [new branch]                gh/mikaylagawarecki/380/head -> origin/gh/mikaylagawarecki/380/head
2025-12-04T08:53:09.5460949Z  * [new branch]                gh/mikaylagawarecki/380/orig -> origin/gh/mikaylagawarecki/380/orig
2025-12-04T08:53:09.5461041Z  * [new branch]                gh/mikaylagawarecki/381/base -> origin/gh/mikaylagawarecki/381/base
2025-12-04T08:53:09.5461171Z  * [new branch]                gh/mikaylagawarecki/381/head -> origin/gh/mikaylagawarecki/381/head
2025-12-04T08:53:09.5461300Z  * [new branch]                gh/mikaylagawarecki/381/orig -> origin/gh/mikaylagawarecki/381/orig
2025-12-04T08:53:09.5461392Z  * [new branch]                gh/mikaylagawarecki/382/base -> origin/gh/mikaylagawarecki/382/base
2025-12-04T08:53:09.5461486Z  * [new branch]                gh/mikaylagawarecki/382/head -> origin/gh/mikaylagawarecki/382/head
2025-12-04T08:53:09.5461577Z  * [new branch]                gh/mikaylagawarecki/382/orig -> origin/gh/mikaylagawarecki/382/orig
2025-12-04T08:53:09.5461669Z  * [new branch]                gh/mikaylagawarecki/383/base -> origin/gh/mikaylagawarecki/383/base
2025-12-04T08:53:09.5461765Z  * [new branch]                gh/mikaylagawarecki/383/head -> origin/gh/mikaylagawarecki/383/head
2025-12-04T08:53:09.5461857Z  * [new branch]                gh/mikaylagawarecki/383/orig -> origin/gh/mikaylagawarecki/383/orig
2025-12-04T08:53:09.5461952Z  * [new branch]                gh/mikaylagawarecki/384/base -> origin/gh/mikaylagawarecki/384/base
2025-12-04T08:53:09.5462045Z  * [new branch]                gh/mikaylagawarecki/384/head -> origin/gh/mikaylagawarecki/384/head
2025-12-04T08:53:09.5462137Z  * [new branch]                gh/mikaylagawarecki/384/orig -> origin/gh/mikaylagawarecki/384/orig
2025-12-04T08:53:09.5462229Z  * [new branch]                gh/mikaylagawarecki/385/base -> origin/gh/mikaylagawarecki/385/base
2025-12-04T08:53:09.5462321Z  * [new branch]                gh/mikaylagawarecki/385/head -> origin/gh/mikaylagawarecki/385/head
2025-12-04T08:53:09.5462417Z  * [new branch]                gh/mikaylagawarecki/385/orig -> origin/gh/mikaylagawarecki/385/orig
2025-12-04T08:53:09.5462511Z  * [new branch]                gh/mikaylagawarecki/386/base -> origin/gh/mikaylagawarecki/386/base
2025-12-04T08:53:09.5462603Z  * [new branch]                gh/mikaylagawarecki/386/head -> origin/gh/mikaylagawarecki/386/head
2025-12-04T08:53:09.5462697Z  * [new branch]                gh/mikaylagawarecki/386/orig -> origin/gh/mikaylagawarecki/386/orig
2025-12-04T08:53:09.5462794Z  * [new branch]                gh/mikaylagawarecki/387/base -> origin/gh/mikaylagawarecki/387/base
2025-12-04T08:53:09.5462885Z  * [new branch]                gh/mikaylagawarecki/387/head -> origin/gh/mikaylagawarecki/387/head
2025-12-04T08:53:09.5462976Z  * [new branch]                gh/mikaylagawarecki/387/orig -> origin/gh/mikaylagawarecki/387/orig
2025-12-04T08:53:09.5463071Z  * [new branch]                gh/mikaylagawarecki/388/base -> origin/gh/mikaylagawarecki/388/base
2025-12-04T08:53:09.5463162Z  * [new branch]                gh/mikaylagawarecki/388/head -> origin/gh/mikaylagawarecki/388/head
2025-12-04T08:53:09.5463255Z  * [new branch]                gh/mikaylagawarecki/388/orig -> origin/gh/mikaylagawarecki/388/orig
2025-12-04T08:53:09.5463347Z  * [new branch]                gh/mikaylagawarecki/389/base -> origin/gh/mikaylagawarecki/389/base
2025-12-04T08:53:09.5463442Z  * [new branch]                gh/mikaylagawarecki/389/head -> origin/gh/mikaylagawarecki/389/head
2025-12-04T08:53:09.5463537Z  * [new branch]                gh/mikaylagawarecki/389/orig -> origin/gh/mikaylagawarecki/389/orig
2025-12-04T08:53:09.5463628Z  * [new branch]                gh/mikaylagawarecki/390/base -> origin/gh/mikaylagawarecki/390/base
2025-12-04T08:53:09.5463720Z  * [new branch]                gh/mikaylagawarecki/390/head -> origin/gh/mikaylagawarecki/390/head
2025-12-04T08:53:09.5463813Z  * [new branch]                gh/mikaylagawarecki/390/orig -> origin/gh/mikaylagawarecki/390/orig
2025-12-04T08:53:09.5463905Z  * [new branch]                gh/mikaylagawarecki/391/base -> origin/gh/mikaylagawarecki/391/base
2025-12-04T08:53:09.5463996Z  * [new branch]                gh/mikaylagawarecki/391/head -> origin/gh/mikaylagawarecki/391/head
2025-12-04T08:53:09.5464090Z  * [new branch]                gh/mikaylagawarecki/391/orig -> origin/gh/mikaylagawarecki/391/orig
2025-12-04T08:53:09.5464212Z  * [new branch]                gh/mikaylagawarecki/392/base -> origin/gh/mikaylagawarecki/392/base
2025-12-04T08:53:09.5464331Z  * [new branch]                gh/mikaylagawarecki/392/head -> origin/gh/mikaylagawarecki/392/head
2025-12-04T08:53:09.5464424Z  * [new branch]                gh/mikaylagawarecki/392/orig -> origin/gh/mikaylagawarecki/392/orig
2025-12-04T08:53:09.5464493Z  * [new branch]                gh/mlazos/41/base       -> origin/gh/mlazos/41/base
2025-12-04T08:53:09.5464562Z  * [new branch]                gh/mlazos/41/head       -> origin/gh/mlazos/41/head
2025-12-04T08:53:09.5464632Z  * [new branch]                gh/mlazos/41/orig       -> origin/gh/mlazos/41/orig
2025-12-04T08:53:09.5464699Z  * [new branch]                gh/mlazos/42/base       -> origin/gh/mlazos/42/base
2025-12-04T08:53:09.5464767Z  * [new branch]                gh/mlazos/42/head       -> origin/gh/mlazos/42/head
2025-12-04T08:53:09.5464834Z  * [new branch]                gh/mlazos/42/orig       -> origin/gh/mlazos/42/orig
2025-12-04T08:53:09.5464901Z  * [new branch]                gh/mlazos/43/base       -> origin/gh/mlazos/43/base
2025-12-04T08:53:09.5464971Z  * [new branch]                gh/mlazos/43/head       -> origin/gh/mlazos/43/head
2025-12-04T08:53:09.5465037Z  * [new branch]                gh/mlazos/43/orig       -> origin/gh/mlazos/43/orig
2025-12-04T08:53:09.5465105Z  * [new branch]                gh/mlazos/44/base       -> origin/gh/mlazos/44/base
2025-12-04T08:53:09.5465173Z  * [new branch]                gh/mlazos/44/head       -> origin/gh/mlazos/44/head
2025-12-04T08:53:09.5465239Z  * [new branch]                gh/mlazos/44/orig       -> origin/gh/mlazos/44/orig
2025-12-04T08:53:09.5465306Z  * [new branch]                gh/mlazos/47/base       -> origin/gh/mlazos/47/base
2025-12-04T08:53:09.5465374Z  * [new branch]                gh/mlazos/47/head       -> origin/gh/mlazos/47/head
2025-12-04T08:53:09.5465440Z  * [new branch]                gh/mlazos/47/orig       -> origin/gh/mlazos/47/orig
2025-12-04T08:53:09.5465508Z  * [new branch]                gh/mlazos/48/base       -> origin/gh/mlazos/48/base
2025-12-04T08:53:09.5465576Z  * [new branch]                gh/mlazos/48/head       -> origin/gh/mlazos/48/head
2025-12-04T08:53:09.5465645Z  * [new branch]                gh/mlazos/48/orig       -> origin/gh/mlazos/48/orig
2025-12-04T08:53:09.5465712Z  * [new branch]                gh/mlazos/49/base       -> origin/gh/mlazos/49/base
2025-12-04T08:53:09.5465779Z  * [new branch]                gh/mlazos/49/head       -> origin/gh/mlazos/49/head
2025-12-04T08:53:09.5465846Z  * [new branch]                gh/mlazos/49/orig       -> origin/gh/mlazos/49/orig
2025-12-04T08:53:09.5465912Z  * [new branch]                gh/mlazos/50/base       -> origin/gh/mlazos/50/base
2025-12-04T08:53:09.5465980Z  * [new branch]                gh/mlazos/50/head       -> origin/gh/mlazos/50/head
2025-12-04T08:53:09.5466046Z  * [new branch]                gh/mlazos/50/orig       -> origin/gh/mlazos/50/orig
2025-12-04T08:53:09.5466114Z  * [new branch]                gh/mlazos/51/base       -> origin/gh/mlazos/51/base
2025-12-04T08:53:09.5466181Z  * [new branch]                gh/mlazos/51/head       -> origin/gh/mlazos/51/head
2025-12-04T08:53:09.5466248Z  * [new branch]                gh/mlazos/51/orig       -> origin/gh/mlazos/51/orig
2025-12-04T08:53:09.5466318Z  * [new branch]                gh/mlazos/52/base       -> origin/gh/mlazos/52/base
2025-12-04T08:53:09.5466384Z  * [new branch]                gh/mlazos/52/head       -> origin/gh/mlazos/52/head
2025-12-04T08:53:09.5466451Z  * [new branch]                gh/mlazos/52/orig       -> origin/gh/mlazos/52/orig
2025-12-04T08:53:09.5466521Z  * [new branch]                gh/mlazos/53/base       -> origin/gh/mlazos/53/base
2025-12-04T08:53:09.5466587Z  * [new branch]                gh/mlazos/53/head       -> origin/gh/mlazos/53/head
2025-12-04T08:53:09.5466652Z  * [new branch]                gh/mlazos/53/orig       -> origin/gh/mlazos/53/orig
2025-12-04T08:53:09.5466743Z  * [new branch]                gh/mlazos/54/base       -> origin/gh/mlazos/54/base
2025-12-04T08:53:09.5466810Z  * [new branch]                gh/mlazos/54/head       -> origin/gh/mlazos/54/head
2025-12-04T08:53:09.5466906Z  * [new branch]                gh/mlazos/54/orig       -> origin/gh/mlazos/54/orig
2025-12-04T08:53:09.5466972Z  * [new branch]                gh/mlazos/55/base       -> origin/gh/mlazos/55/base
2025-12-04T08:53:09.5467037Z  * [new branch]                gh/mlazos/55/head       -> origin/gh/mlazos/55/head
2025-12-04T08:53:09.5467102Z  * [new branch]                gh/mlazos/55/orig       -> origin/gh/mlazos/55/orig
2025-12-04T08:53:09.5467171Z  * [new branch]                gh/mlazos/56/base       -> origin/gh/mlazos/56/base
2025-12-04T08:53:09.5467237Z  * [new branch]                gh/mlazos/56/head       -> origin/gh/mlazos/56/head
2025-12-04T08:53:09.5467304Z  * [new branch]                gh/mlazos/56/orig       -> origin/gh/mlazos/56/orig
2025-12-04T08:53:09.5467375Z  * [new branch]                gh/mlazos/57/base       -> origin/gh/mlazos/57/base
2025-12-04T08:53:09.5467443Z  * [new branch]                gh/mlazos/57/head       -> origin/gh/mlazos/57/head
2025-12-04T08:53:09.5467511Z  * [new branch]                gh/mlazos/57/orig       -> origin/gh/mlazos/57/orig
2025-12-04T08:53:09.5467580Z  * [new branch]                gh/mlazos/58/base       -> origin/gh/mlazos/58/base
2025-12-04T08:53:09.5467646Z  * [new branch]                gh/mlazos/58/head       -> origin/gh/mlazos/58/head
2025-12-04T08:53:09.5467712Z  * [new branch]                gh/mlazos/58/orig       -> origin/gh/mlazos/58/orig
2025-12-04T08:53:09.5467781Z  * [new branch]                gh/mlazos/59/base       -> origin/gh/mlazos/59/base
2025-12-04T08:53:09.5467847Z  * [new branch]                gh/mlazos/59/head       -> origin/gh/mlazos/59/head
2025-12-04T08:53:09.5467914Z  * [new branch]                gh/mlazos/59/orig       -> origin/gh/mlazos/59/orig
2025-12-04T08:53:09.5467980Z  * [new branch]                gh/mlazos/60/base       -> origin/gh/mlazos/60/base
2025-12-04T08:53:09.5468047Z  * [new branch]                gh/mlazos/60/head       -> origin/gh/mlazos/60/head
2025-12-04T08:53:09.5468116Z  * [new branch]                gh/mlazos/60/orig       -> origin/gh/mlazos/60/orig
2025-12-04T08:53:09.5468183Z  * [new branch]                gh/mlazos/61/base       -> origin/gh/mlazos/61/base
2025-12-04T08:53:09.5468250Z  * [new branch]                gh/mlazos/61/head       -> origin/gh/mlazos/61/head
2025-12-04T08:53:09.5468320Z  * [new branch]                gh/mlazos/61/orig       -> origin/gh/mlazos/61/orig
2025-12-04T08:53:09.5468387Z  * [new branch]                gh/mlazos/62/base       -> origin/gh/mlazos/62/base
2025-12-04T08:53:09.5468453Z  * [new branch]                gh/mlazos/62/head       -> origin/gh/mlazos/62/head
2025-12-04T08:53:09.5468521Z  * [new branch]                gh/mlazos/62/orig       -> origin/gh/mlazos/62/orig
2025-12-04T08:53:09.5468587Z  * [new branch]                gh/mlazos/63/base       -> origin/gh/mlazos/63/base
2025-12-04T08:53:09.5468659Z  * [new branch]                gh/mlazos/63/head       -> origin/gh/mlazos/63/head
2025-12-04T08:53:09.5468726Z  * [new branch]                gh/mlazos/63/orig       -> origin/gh/mlazos/63/orig
2025-12-04T08:53:09.5468791Z  * [new branch]                gh/mlazos/64/base       -> origin/gh/mlazos/64/base
2025-12-04T08:53:09.5468857Z  * [new branch]                gh/mlazos/64/head       -> origin/gh/mlazos/64/head
2025-12-04T08:53:09.5468923Z  * [new branch]                gh/mlazos/64/orig       -> origin/gh/mlazos/64/orig
2025-12-04T08:53:09.5468989Z  * [new branch]                gh/mlazos/65/base       -> origin/gh/mlazos/65/base
2025-12-04T08:53:09.5469054Z  * [new branch]                gh/mlazos/65/head       -> origin/gh/mlazos/65/head
2025-12-04T08:53:09.5469121Z  * [new branch]                gh/mlazos/65/orig       -> origin/gh/mlazos/65/orig
2025-12-04T08:53:09.5469187Z  * [new branch]                gh/mlazos/66/base       -> origin/gh/mlazos/66/base
2025-12-04T08:53:09.5469280Z  * [new branch]                gh/mlazos/66/head       -> origin/gh/mlazos/66/head
2025-12-04T08:53:09.5469372Z  * [new branch]                gh/mlazos/66/orig       -> origin/gh/mlazos/66/orig
2025-12-04T08:53:09.5469439Z  * [new branch]                gh/mlazos/67/base       -> origin/gh/mlazos/67/base
2025-12-04T08:53:09.5469505Z  * [new branch]                gh/mlazos/67/head       -> origin/gh/mlazos/67/head
2025-12-04T08:53:09.5469570Z  * [new branch]                gh/mlazos/67/orig       -> origin/gh/mlazos/67/orig
2025-12-04T08:53:09.5469635Z  * [new branch]                gh/mlazos/68/base       -> origin/gh/mlazos/68/base
2025-12-04T08:53:09.5469703Z  * [new branch]                gh/mlazos/68/head       -> origin/gh/mlazos/68/head
2025-12-04T08:53:09.5469768Z  * [new branch]                gh/mlazos/68/orig       -> origin/gh/mlazos/68/orig
2025-12-04T08:53:09.5469833Z  * [new branch]                gh/mlazos/69/base       -> origin/gh/mlazos/69/base
2025-12-04T08:53:09.5469903Z  * [new branch]                gh/mlazos/69/head       -> origin/gh/mlazos/69/head
2025-12-04T08:53:09.5469969Z  * [new branch]                gh/mlazos/69/orig       -> origin/gh/mlazos/69/orig
2025-12-04T08:53:09.5470035Z  * [new branch]                gh/mlazos/70/base       -> origin/gh/mlazos/70/base
2025-12-04T08:53:09.5470102Z  * [new branch]                gh/mlazos/70/head       -> origin/gh/mlazos/70/head
2025-12-04T08:53:09.5470167Z  * [new branch]                gh/mlazos/70/orig       -> origin/gh/mlazos/70/orig
2025-12-04T08:53:09.5470232Z  * [new branch]                gh/mlazos/71/base       -> origin/gh/mlazos/71/base
2025-12-04T08:53:09.5470298Z  * [new branch]                gh/mlazos/71/head       -> origin/gh/mlazos/71/head
2025-12-04T08:53:09.5470363Z  * [new branch]                gh/mlazos/71/orig       -> origin/gh/mlazos/71/orig
2025-12-04T08:53:09.5470461Z  * [new branch]                gh/mlazos/72/base       -> origin/gh/mlazos/72/base
2025-12-04T08:53:09.5470530Z  * [new branch]                gh/mlazos/72/head       -> origin/gh/mlazos/72/head
2025-12-04T08:53:09.5470595Z  * [new branch]                gh/mlazos/72/orig       -> origin/gh/mlazos/72/orig
2025-12-04T08:53:09.5470662Z  * [new branch]                gh/mlazos/73/base       -> origin/gh/mlazos/73/base
2025-12-04T08:53:09.5470729Z  * [new branch]                gh/mlazos/73/head       -> origin/gh/mlazos/73/head
2025-12-04T08:53:09.5470794Z  * [new branch]                gh/mlazos/73/orig       -> origin/gh/mlazos/73/orig
2025-12-04T08:53:09.5470862Z  * [new branch]                gh/mrmiywj/1/base       -> origin/gh/mrmiywj/1/base
2025-12-04T08:53:09.5470930Z  * [new branch]                gh/mrmiywj/1/head       -> origin/gh/mrmiywj/1/head
2025-12-04T08:53:09.5471004Z  * [new branch]                gh/muchulee8/73/base    -> origin/gh/muchulee8/73/base
2025-12-04T08:53:09.5471076Z  * [new branch]                gh/muchulee8/73/head    -> origin/gh/muchulee8/73/head
2025-12-04T08:53:09.5471152Z  * [new branch]                gh/muchulee8/73/orig    -> origin/gh/muchulee8/73/orig
2025-12-04T08:53:09.5471238Z  * [new branch]                gh/naveenthangudu/1/base -> origin/gh/naveenthangudu/1/base
2025-12-04T08:53:09.5471322Z  * [new branch]                gh/naveenthangudu/1/head -> origin/gh/naveenthangudu/1/head
2025-12-04T08:53:09.5471403Z  * [new branch]                gh/naveenthangudu/1/orig -> origin/gh/naveenthangudu/1/orig
2025-12-04T08:53:09.5471482Z  * [new branch]                gh/naveenthangudu/2/base -> origin/gh/naveenthangudu/2/base
2025-12-04T08:53:09.5471562Z  * [new branch]                gh/naveenthangudu/2/head -> origin/gh/naveenthangudu/2/head
2025-12-04T08:53:09.5471640Z  * [new branch]                gh/naveenthangudu/2/orig -> origin/gh/naveenthangudu/2/orig
2025-12-04T08:53:09.5471718Z  * [new branch]                gh/naveenthangudu/3/base -> origin/gh/naveenthangudu/3/base
2025-12-04T08:53:09.5471798Z  * [new branch]                gh/naveenthangudu/3/head -> origin/gh/naveenthangudu/3/head
2025-12-04T08:53:09.5471923Z  * [new branch]                gh/naveenthangudu/3/orig -> origin/gh/naveenthangudu/3/orig
2025-12-04T08:53:09.5472044Z  * [new branch]                gh/naveenthangudu/4/base -> origin/gh/naveenthangudu/4/base
2025-12-04T08:53:09.5472126Z  * [new branch]                gh/naveenthangudu/4/head -> origin/gh/naveenthangudu/4/head
2025-12-04T08:53:09.5472206Z  * [new branch]                gh/naveenthangudu/4/orig -> origin/gh/naveenthangudu/4/orig
2025-12-04T08:53:09.5472285Z  * [new branch]                gh/naveenthangudu/5/base -> origin/gh/naveenthangudu/5/base
2025-12-04T08:53:09.5472366Z  * [new branch]                gh/naveenthangudu/5/head -> origin/gh/naveenthangudu/5/head
2025-12-04T08:53:09.5472446Z  * [new branch]                gh/naveenthangudu/5/orig -> origin/gh/naveenthangudu/5/orig
2025-12-04T08:53:09.5472525Z  * [new branch]                gh/naveenthangudu/6/base -> origin/gh/naveenthangudu/6/base
2025-12-04T08:53:09.5472608Z  * [new branch]                gh/naveenthangudu/6/head -> origin/gh/naveenthangudu/6/head
2025-12-04T08:53:09.5472688Z  * [new branch]                gh/naveenthangudu/6/orig -> origin/gh/naveenthangudu/6/orig
2025-12-04T08:53:09.5472771Z  * [new branch]                gh/naveenthangudu/7/base -> origin/gh/naveenthangudu/7/base
2025-12-04T08:53:09.5472852Z  * [new branch]                gh/naveenthangudu/7/head -> origin/gh/naveenthangudu/7/head
2025-12-04T08:53:09.5472932Z  * [new branch]                gh/naveenthangudu/7/orig -> origin/gh/naveenthangudu/7/orig
2025-12-04T08:53:09.5473015Z  * [new branch]                gh/naveenthangudu/8/base -> origin/gh/naveenthangudu/8/base
2025-12-04T08:53:09.5473095Z  * [new branch]                gh/naveenthangudu/8/head -> origin/gh/naveenthangudu/8/head
2025-12-04T08:53:09.5473174Z  * [new branch]                gh/naveenthangudu/8/orig -> origin/gh/naveenthangudu/8/orig
2025-12-04T08:53:09.5473255Z  * [new branch]                gh/naveenthangudu/9/base -> origin/gh/naveenthangudu/9/base
2025-12-04T08:53:09.5473337Z  * [new branch]                gh/naveenthangudu/9/head -> origin/gh/naveenthangudu/9/head
2025-12-04T08:53:09.5473417Z  * [new branch]                gh/naveenthangudu/9/orig -> origin/gh/naveenthangudu/9/orig
2025-12-04T08:53:09.5473491Z  * [new branch]                gh/nikitaved/1/base     -> origin/gh/nikitaved/1/base
2025-12-04T08:53:09.5473564Z  * [new branch]                gh/nikitaved/1/head     -> origin/gh/nikitaved/1/head
2025-12-04T08:53:09.5473635Z  * [new branch]                gh/nikitaved/1/orig     -> origin/gh/nikitaved/1/orig
2025-12-04T08:53:09.5473711Z  * [new branch]                gh/nikitaved/10/base    -> origin/gh/nikitaved/10/base
2025-12-04T08:53:09.5473784Z  * [new branch]                gh/nikitaved/10/head    -> origin/gh/nikitaved/10/head
2025-12-04T08:53:09.5473856Z  * [new branch]                gh/nikitaved/10/orig    -> origin/gh/nikitaved/10/orig
2025-12-04T08:53:09.5473931Z  * [new branch]                gh/nikitaved/11/base    -> origin/gh/nikitaved/11/base
2025-12-04T08:53:09.5474004Z  * [new branch]                gh/nikitaved/11/head    -> origin/gh/nikitaved/11/head
2025-12-04T08:53:09.5474077Z  * [new branch]                gh/nikitaved/11/orig    -> origin/gh/nikitaved/11/orig
2025-12-04T08:53:09.5474153Z  * [new branch]                gh/nikitaved/12/base    -> origin/gh/nikitaved/12/base
2025-12-04T08:53:09.5474224Z  * [new branch]                gh/nikitaved/12/head    -> origin/gh/nikitaved/12/head
2025-12-04T08:53:09.5474298Z  * [new branch]                gh/nikitaved/12/orig    -> origin/gh/nikitaved/12/orig
2025-12-04T08:53:09.5474370Z  * [new branch]                gh/nikitaved/13/base    -> origin/gh/nikitaved/13/base
2025-12-04T08:53:09.5474442Z  * [new branch]                gh/nikitaved/13/head    -> origin/gh/nikitaved/13/head
2025-12-04T08:53:09.5474514Z  * [new branch]                gh/nikitaved/13/orig    -> origin/gh/nikitaved/13/orig
2025-12-04T08:53:09.5474586Z  * [new branch]                gh/nikitaved/14/base    -> origin/gh/nikitaved/14/base
2025-12-04T08:53:09.5474686Z  * [new branch]                gh/nikitaved/14/head    -> origin/gh/nikitaved/14/head
2025-12-04T08:53:09.5474789Z  * [new branch]                gh/nikitaved/14/orig    -> origin/gh/nikitaved/14/orig
2025-12-04T08:53:09.5474861Z  * [new branch]                gh/nikitaved/15/base    -> origin/gh/nikitaved/15/base
2025-12-04T08:53:09.5474933Z  * [new branch]                gh/nikitaved/15/head    -> origin/gh/nikitaved/15/head
2025-12-04T08:53:09.5475007Z  * [new branch]                gh/nikitaved/15/orig    -> origin/gh/nikitaved/15/orig
2025-12-04T08:53:09.5475079Z  * [new branch]                gh/nikitaved/16/base    -> origin/gh/nikitaved/16/base
2025-12-04T08:53:09.5475151Z  * [new branch]                gh/nikitaved/16/head    -> origin/gh/nikitaved/16/head
2025-12-04T08:53:09.5475224Z  * [new branch]                gh/nikitaved/16/orig    -> origin/gh/nikitaved/16/orig
2025-12-04T08:53:09.5475296Z  * [new branch]                gh/nikitaved/2/base     -> origin/gh/nikitaved/2/base
2025-12-04T08:53:09.5475370Z  * [new branch]                gh/nikitaved/2/head     -> origin/gh/nikitaved/2/head
2025-12-04T08:53:09.5475445Z  * [new branch]                gh/nikitaved/2/orig     -> origin/gh/nikitaved/2/orig
2025-12-04T08:53:09.5475515Z  * [new branch]                gh/nikitaved/4/base     -> origin/gh/nikitaved/4/base
2025-12-04T08:53:09.5475585Z  * [new branch]                gh/nikitaved/4/head     -> origin/gh/nikitaved/4/head
2025-12-04T08:53:09.5475659Z  * [new branch]                gh/nikitaved/4/orig     -> origin/gh/nikitaved/4/orig
2025-12-04T08:53:09.5475730Z  * [new branch]                gh/nikitaved/5/base     -> origin/gh/nikitaved/5/base
2025-12-04T08:53:09.5475801Z  * [new branch]                gh/nikitaved/5/head     -> origin/gh/nikitaved/5/head
2025-12-04T08:53:09.5475871Z  * [new branch]                gh/nikitaved/5/orig     -> origin/gh/nikitaved/5/orig
2025-12-04T08:53:09.5475942Z  * [new branch]                gh/nikitaved/6/base     -> origin/gh/nikitaved/6/base
2025-12-04T08:53:09.5476016Z  * [new branch]                gh/nikitaved/6/head     -> origin/gh/nikitaved/6/head
2025-12-04T08:53:09.5476087Z  * [new branch]                gh/nikitaved/6/orig     -> origin/gh/nikitaved/6/orig
2025-12-04T08:53:09.5476156Z  * [new branch]                gh/nikitaved/8/base     -> origin/gh/nikitaved/8/base
2025-12-04T08:53:09.5476229Z  * [new branch]                gh/nikitaved/8/head     -> origin/gh/nikitaved/8/head
2025-12-04T08:53:09.5476298Z  * [new branch]                gh/nikitaved/8/orig     -> origin/gh/nikitaved/8/orig
2025-12-04T08:53:09.5476369Z  * [new branch]                gh/nikitaved/9/base     -> origin/gh/nikitaved/9/base
2025-12-04T08:53:09.5476441Z  * [new branch]                gh/nikitaved/9/head     -> origin/gh/nikitaved/9/head
2025-12-04T08:53:09.5476512Z  * [new branch]                gh/nikitaved/9/orig     -> origin/gh/nikitaved/9/orig
2025-12-04T08:53:09.5476581Z  * [new branch]                gh/oulgen/10/base       -> origin/gh/oulgen/10/base
2025-12-04T08:53:09.5476653Z  * [new branch]                gh/oulgen/10/head       -> origin/gh/oulgen/10/head
2025-12-04T08:53:09.5476724Z  * [new branch]                gh/oulgen/10/orig       -> origin/gh/oulgen/10/orig
2025-12-04T08:53:09.5476790Z  * [new branch]                gh/oulgen/11/base       -> origin/gh/oulgen/11/base
2025-12-04T08:53:09.5476860Z  * [new branch]                gh/oulgen/11/head       -> origin/gh/oulgen/11/head
2025-12-04T08:53:09.5476926Z  * [new branch]                gh/oulgen/11/orig       -> origin/gh/oulgen/11/orig
2025-12-04T08:53:09.5476991Z  * [new branch]                gh/oulgen/12/base       -> origin/gh/oulgen/12/base
2025-12-04T08:53:09.5477059Z  * [new branch]                gh/oulgen/12/head       -> origin/gh/oulgen/12/head
2025-12-04T08:53:09.5477126Z  * [new branch]                gh/oulgen/12/orig       -> origin/gh/oulgen/12/orig
2025-12-04T08:53:09.5477192Z  * [new branch]                gh/oulgen/13/base       -> origin/gh/oulgen/13/base
2025-12-04T08:53:09.5477292Z  * [new branch]                gh/oulgen/13/head       -> origin/gh/oulgen/13/head
2025-12-04T08:53:09.5477379Z  * [new branch]                gh/oulgen/13/orig       -> origin/gh/oulgen/13/orig
2025-12-04T08:53:09.5477446Z  * [new branch]                gh/oulgen/14/base       -> origin/gh/oulgen/14/base
2025-12-04T08:53:09.5477515Z  * [new branch]                gh/oulgen/14/head       -> origin/gh/oulgen/14/head
2025-12-04T08:53:09.5477581Z  * [new branch]                gh/oulgen/14/orig       -> origin/gh/oulgen/14/orig
2025-12-04T08:53:09.5477649Z  * [new branch]                gh/oulgen/15/base       -> origin/gh/oulgen/15/base
2025-12-04T08:53:09.5477715Z  * [new branch]                gh/oulgen/15/head       -> origin/gh/oulgen/15/head
2025-12-04T08:53:09.5477780Z  * [new branch]                gh/oulgen/15/orig       -> origin/gh/oulgen/15/orig
2025-12-04T08:53:09.5477848Z  * [new branch]                gh/oulgen/16/base       -> origin/gh/oulgen/16/base
2025-12-04T08:53:09.5477916Z  * [new branch]                gh/oulgen/16/head       -> origin/gh/oulgen/16/head
2025-12-04T08:53:09.5477985Z  * [new branch]                gh/oulgen/16/orig       -> origin/gh/oulgen/16/orig
2025-12-04T08:53:09.5478053Z  * [new branch]                gh/oulgen/17/base       -> origin/gh/oulgen/17/base
2025-12-04T08:53:09.5478118Z  * [new branch]                gh/oulgen/17/head       -> origin/gh/oulgen/17/head
2025-12-04T08:53:09.5478185Z  * [new branch]                gh/oulgen/17/orig       -> origin/gh/oulgen/17/orig
2025-12-04T08:53:09.5478253Z  * [new branch]                gh/oulgen/18/base       -> origin/gh/oulgen/18/base
2025-12-04T08:53:09.5478319Z  * [new branch]                gh/oulgen/18/head       -> origin/gh/oulgen/18/head
2025-12-04T08:53:09.5478384Z  * [new branch]                gh/oulgen/18/orig       -> origin/gh/oulgen/18/orig
2025-12-04T08:53:09.5478453Z  * [new branch]                gh/oulgen/19/base       -> origin/gh/oulgen/19/base
2025-12-04T08:53:09.5478520Z  * [new branch]                gh/oulgen/19/head       -> origin/gh/oulgen/19/head
2025-12-04T08:53:09.5478588Z  * [new branch]                gh/oulgen/19/orig       -> origin/gh/oulgen/19/orig
2025-12-04T08:53:09.5478656Z  * [new branch]                gh/oulgen/20/base       -> origin/gh/oulgen/20/base
2025-12-04T08:53:09.5478720Z  * [new branch]                gh/oulgen/20/head       -> origin/gh/oulgen/20/head
2025-12-04T08:53:09.5478785Z  * [new branch]                gh/oulgen/20/orig       -> origin/gh/oulgen/20/orig
2025-12-04T08:53:09.5478852Z  * [new branch]                gh/oulgen/21/base       -> origin/gh/oulgen/21/base
2025-12-04T08:53:09.5478918Z  * [new branch]                gh/oulgen/21/head       -> origin/gh/oulgen/21/head
2025-12-04T08:53:09.5478986Z  * [new branch]                gh/oulgen/21/orig       -> origin/gh/oulgen/21/orig
2025-12-04T08:53:09.5479055Z  * [new branch]                gh/oulgen/22/base       -> origin/gh/oulgen/22/base
2025-12-04T08:53:09.5479123Z  * [new branch]                gh/oulgen/22/head       -> origin/gh/oulgen/22/head
2025-12-04T08:53:09.5479190Z  * [new branch]                gh/oulgen/22/orig       -> origin/gh/oulgen/22/orig
2025-12-04T08:53:09.5479260Z  * [new branch]                gh/oulgen/23/base       -> origin/gh/oulgen/23/base
2025-12-04T08:53:09.5479327Z  * [new branch]                gh/oulgen/23/head       -> origin/gh/oulgen/23/head
2025-12-04T08:53:09.5479396Z  * [new branch]                gh/oulgen/23/orig       -> origin/gh/oulgen/23/orig
2025-12-04T08:53:09.5479461Z  * [new branch]                gh/oulgen/24/base       -> origin/gh/oulgen/24/base
2025-12-04T08:53:09.5479527Z  * [new branch]                gh/oulgen/24/head       -> origin/gh/oulgen/24/head
2025-12-04T08:53:09.5479593Z  * [new branch]                gh/oulgen/24/orig       -> origin/gh/oulgen/24/orig
2025-12-04T08:53:09.5479658Z  * [new branch]                gh/oulgen/25/base       -> origin/gh/oulgen/25/base
2025-12-04T08:53:09.5479752Z  * [new branch]                gh/oulgen/25/head       -> origin/gh/oulgen/25/head
2025-12-04T08:53:09.5479818Z  * [new branch]                gh/oulgen/25/orig       -> origin/gh/oulgen/25/orig
2025-12-04T08:53:09.5479910Z  * [new branch]                gh/oulgen/26/base       -> origin/gh/oulgen/26/base
2025-12-04T08:53:09.5479977Z  * [new branch]                gh/oulgen/26/head       -> origin/gh/oulgen/26/head
2025-12-04T08:53:09.5480045Z  * [new branch]                gh/oulgen/26/orig       -> origin/gh/oulgen/26/orig
2025-12-04T08:53:09.5480115Z  * [new branch]                gh/oulgen/4/base        -> origin/gh/oulgen/4/base
2025-12-04T08:53:09.5480183Z  * [new branch]                gh/oulgen/4/head        -> origin/gh/oulgen/4/head
2025-12-04T08:53:09.5480251Z  * [new branch]                gh/oulgen/4/orig        -> origin/gh/oulgen/4/orig
2025-12-04T08:53:09.5480317Z  * [new branch]                gh/oulgen/7/base        -> origin/gh/oulgen/7/base
2025-12-04T08:53:09.5480385Z  * [new branch]                gh/oulgen/7/head        -> origin/gh/oulgen/7/head
2025-12-04T08:53:09.5480490Z  * [new branch]                gh/oulgen/7/orig        -> origin/gh/oulgen/7/orig
2025-12-04T08:53:09.5480558Z  * [new branch]                gh/oulgen/8/base        -> origin/gh/oulgen/8/base
2025-12-04T08:53:09.5480624Z  * [new branch]                gh/oulgen/8/head        -> origin/gh/oulgen/8/head
2025-12-04T08:53:09.5480691Z  * [new branch]                gh/oulgen/8/orig        -> origin/gh/oulgen/8/orig
2025-12-04T08:53:09.5480759Z  * [new branch]                gh/oulgen/9/base        -> origin/gh/oulgen/9/base
2025-12-04T08:53:09.5480823Z  * [new branch]                gh/oulgen/9/head        -> origin/gh/oulgen/9/head
2025-12-04T08:53:09.5480892Z  * [new branch]                gh/oulgen/9/orig        -> origin/gh/oulgen/9/orig
2025-12-04T08:53:09.5480997Z  * [new branch]                gh/patvig/mtia-serialization -> origin/gh/patvig/mtia-serialization
2025-12-04T08:53:09.5481066Z  * [new branch]                gh/pearu/108/base       -> origin/gh/pearu/108/base
2025-12-04T08:53:09.5481135Z  * [new branch]                gh/pearu/108/head       -> origin/gh/pearu/108/head
2025-12-04T08:53:09.5481203Z  * [new branch]                gh/pearu/108/orig       -> origin/gh/pearu/108/orig
2025-12-04T08:53:09.5481274Z  * [new branch]                gh/pearu/109/base       -> origin/gh/pearu/109/base
2025-12-04T08:53:09.5481340Z  * [new branch]                gh/pearu/109/head       -> origin/gh/pearu/109/head
2025-12-04T08:53:09.5481406Z  * [new branch]                gh/pearu/109/orig       -> origin/gh/pearu/109/orig
2025-12-04T08:53:09.5481475Z  * [new branch]                gh/pearu/110/base       -> origin/gh/pearu/110/base
2025-12-04T08:53:09.5481541Z  * [new branch]                gh/pearu/110/head       -> origin/gh/pearu/110/head
2025-12-04T08:53:09.5481608Z  * [new branch]                gh/pearu/110/orig       -> origin/gh/pearu/110/orig
2025-12-04T08:53:09.5481675Z  * [new branch]                gh/pearu/111/base       -> origin/gh/pearu/111/base
2025-12-04T08:53:09.5481745Z  * [new branch]                gh/pearu/111/head       -> origin/gh/pearu/111/head
2025-12-04T08:53:09.5481813Z  * [new branch]                gh/pearu/111/orig       -> origin/gh/pearu/111/orig
2025-12-04T08:53:09.5481883Z  * [new branch]                gh/pearu/112/base       -> origin/gh/pearu/112/base
2025-12-04T08:53:09.5481949Z  * [new branch]                gh/pearu/112/head       -> origin/gh/pearu/112/head
2025-12-04T08:53:09.5482015Z  * [new branch]                gh/pearu/112/orig       -> origin/gh/pearu/112/orig
2025-12-04T08:53:09.5482084Z  * [new branch]                gh/pearu/115/base       -> origin/gh/pearu/115/base
2025-12-04T08:53:09.5482154Z  * [new branch]                gh/pearu/115/head       -> origin/gh/pearu/115/head
2025-12-04T08:53:09.5482220Z  * [new branch]                gh/pearu/115/orig       -> origin/gh/pearu/115/orig
2025-12-04T08:53:09.5482287Z  * [new branch]                gh/pearu/116/base       -> origin/gh/pearu/116/base
2025-12-04T08:53:09.5482398Z  * [new branch]                gh/pearu/116/head       -> origin/gh/pearu/116/head
2025-12-04T08:53:09.5482505Z  * [new branch]                gh/pearu/116/orig       -> origin/gh/pearu/116/orig
2025-12-04T08:53:09.5482575Z  * [new branch]                gh/pearu/117/base       -> origin/gh/pearu/117/base
2025-12-04T08:53:09.5482641Z  * [new branch]                gh/pearu/117/head       -> origin/gh/pearu/117/head
2025-12-04T08:53:09.5482712Z  * [new branch]                gh/pearu/117/orig       -> origin/gh/pearu/117/orig
2025-12-04T08:53:09.5482779Z  * [new branch]                gh/pearu/118/base       -> origin/gh/pearu/118/base
2025-12-04T08:53:09.5482844Z  * [new branch]                gh/pearu/118/head       -> origin/gh/pearu/118/head
2025-12-04T08:53:09.5482910Z  * [new branch]                gh/pearu/118/orig       -> origin/gh/pearu/118/orig
2025-12-04T08:53:09.5482977Z  * [new branch]                gh/pearu/119/base       -> origin/gh/pearu/119/base
2025-12-04T08:53:09.5483045Z  * [new branch]                gh/pearu/119/head       -> origin/gh/pearu/119/head
2025-12-04T08:53:09.5483114Z  * [new branch]                gh/pearu/119/orig       -> origin/gh/pearu/119/orig
2025-12-04T08:53:09.5483180Z  * [new branch]                gh/pearu/139/base       -> origin/gh/pearu/139/base
2025-12-04T08:53:09.5483247Z  * [new branch]                gh/pearu/139/head       -> origin/gh/pearu/139/head
2025-12-04T08:53:09.5483313Z  * [new branch]                gh/pearu/139/orig       -> origin/gh/pearu/139/orig
2025-12-04T08:53:09.5483380Z  * [new branch]                gh/pearu/140/base       -> origin/gh/pearu/140/base
2025-12-04T08:53:09.5483447Z  * [new branch]                gh/pearu/140/head       -> origin/gh/pearu/140/head
2025-12-04T08:53:09.5483516Z  * [new branch]                gh/pearu/140/orig       -> origin/gh/pearu/140/orig
2025-12-04T08:53:09.5483583Z  * [new branch]                gh/pearu/142/base       -> origin/gh/pearu/142/base
2025-12-04T08:53:09.5483651Z  * [new branch]                gh/pearu/142/head       -> origin/gh/pearu/142/head
2025-12-04T08:53:09.5483721Z  * [new branch]                gh/pearu/142/orig       -> origin/gh/pearu/142/orig
2025-12-04T08:53:09.5483789Z  * [new branch]                gh/pearu/143/base       -> origin/gh/pearu/143/base
2025-12-04T08:53:09.5483854Z  * [new branch]                gh/pearu/143/head       -> origin/gh/pearu/143/head
2025-12-04T08:53:09.5483922Z  * [new branch]                gh/pearu/143/orig       -> origin/gh/pearu/143/orig
2025-12-04T08:53:09.5483988Z  * [new branch]                gh/pearu/147/base       -> origin/gh/pearu/147/base
2025-12-04T08:53:09.5484054Z  * [new branch]                gh/pearu/147/head       -> origin/gh/pearu/147/head
2025-12-04T08:53:09.5484121Z  * [new branch]                gh/pearu/147/orig       -> origin/gh/pearu/147/orig
2025-12-04T08:53:09.5484186Z  * [new branch]                gh/pearu/149/base       -> origin/gh/pearu/149/base
2025-12-04T08:53:09.5484255Z  * [new branch]                gh/pearu/149/head       -> origin/gh/pearu/149/head
2025-12-04T08:53:09.5484322Z  * [new branch]                gh/pearu/149/orig       -> origin/gh/pearu/149/orig
2025-12-04T08:53:09.5484389Z  * [new branch]                gh/pearu/150/base       -> origin/gh/pearu/150/base
2025-12-04T08:53:09.5484458Z  * [new branch]                gh/pearu/150/head       -> origin/gh/pearu/150/head
2025-12-04T08:53:09.5484525Z  * [new branch]                gh/pearu/150/orig       -> origin/gh/pearu/150/orig
2025-12-04T08:53:09.5484592Z  * [new branch]                gh/pearu/151/base       -> origin/gh/pearu/151/base
2025-12-04T08:53:09.5484659Z  * [new branch]                gh/pearu/151/head       -> origin/gh/pearu/151/head
2025-12-04T08:53:09.5484726Z  * [new branch]                gh/pearu/151/orig       -> origin/gh/pearu/151/orig
2025-12-04T08:53:09.5484791Z  * [new branch]                gh/pearu/152/base       -> origin/gh/pearu/152/base
2025-12-04T08:53:09.5484891Z  * [new branch]                gh/pearu/152/head       -> origin/gh/pearu/152/head
2025-12-04T08:53:09.5484958Z  * [new branch]                gh/pearu/152/orig       -> origin/gh/pearu/152/orig
2025-12-04T08:53:09.5485047Z  * [new branch]                gh/pearu/153/base       -> origin/gh/pearu/153/base
2025-12-04T08:53:09.5485116Z  * [new branch]                gh/pearu/153/head       -> origin/gh/pearu/153/head
2025-12-04T08:53:09.5485181Z  * [new branch]                gh/pearu/153/orig       -> origin/gh/pearu/153/orig
2025-12-04T08:53:09.5485248Z  * [new branch]                gh/pearu/154/base       -> origin/gh/pearu/154/base
2025-12-04T08:53:09.5485314Z  * [new branch]                gh/pearu/154/head       -> origin/gh/pearu/154/head
2025-12-04T08:53:09.5485380Z  * [new branch]                gh/pearu/154/orig       -> origin/gh/pearu/154/orig
2025-12-04T08:53:09.5485446Z  * [new branch]                gh/pearu/155/base       -> origin/gh/pearu/155/base
2025-12-04T08:53:09.5485512Z  * [new branch]                gh/pearu/155/head       -> origin/gh/pearu/155/head
2025-12-04T08:53:09.5485581Z  * [new branch]                gh/pearu/155/orig       -> origin/gh/pearu/155/orig
2025-12-04T08:53:09.5485648Z  * [new branch]                gh/pearu/156/base       -> origin/gh/pearu/156/base
2025-12-04T08:53:09.5485716Z  * [new branch]                gh/pearu/156/head       -> origin/gh/pearu/156/head
2025-12-04T08:53:09.5485783Z  * [new branch]                gh/pearu/156/orig       -> origin/gh/pearu/156/orig
2025-12-04T08:53:09.5485853Z  * [new branch]                gh/pearu/56/base        -> origin/gh/pearu/56/base
2025-12-04T08:53:09.5485921Z  * [new branch]                gh/pearu/56/head        -> origin/gh/pearu/56/head
2025-12-04T08:53:09.5485988Z  * [new branch]                gh/pearu/56/orig        -> origin/gh/pearu/56/orig
2025-12-04T08:53:09.5486059Z  * [new branch]                gh/pearu/97/base        -> origin/gh/pearu/97/base
2025-12-04T08:53:09.5486126Z  * [new branch]                gh/pearu/97/head        -> origin/gh/pearu/97/head
2025-12-04T08:53:09.5486194Z  * [new branch]                gh/pearu/97/orig        -> origin/gh/pearu/97/orig
2025-12-04T08:53:09.5486276Z  * [new branch]                gh/pianpwk/21/base      -> origin/gh/pianpwk/21/base
2025-12-04T08:53:09.5486352Z  * [new branch]                gh/pianpwk/21/head      -> origin/gh/pianpwk/21/head
2025-12-04T08:53:09.5486425Z  * [new branch]                gh/pianpwk/28/base      -> origin/gh/pianpwk/28/base
2025-12-04T08:53:09.5486500Z  * [new branch]                gh/pianpwk/28/head      -> origin/gh/pianpwk/28/head
2025-12-04T08:53:09.5486570Z  * [new branch]                gh/pianpwk/28/orig      -> origin/gh/pianpwk/28/orig
2025-12-04T08:53:09.5486640Z  * [new branch]                gh/pianpwk/29/base      -> origin/gh/pianpwk/29/base
2025-12-04T08:53:09.5486713Z  * [new branch]                gh/pianpwk/29/head      -> origin/gh/pianpwk/29/head
2025-12-04T08:53:09.5486783Z  * [new branch]                gh/pianpwk/29/orig      -> origin/gh/pianpwk/29/orig
2025-12-04T08:53:09.5486855Z  * [new branch]                gh/pianpwk/30/base      -> origin/gh/pianpwk/30/base
2025-12-04T08:53:09.5486931Z  * [new branch]                gh/pianpwk/30/head      -> origin/gh/pianpwk/30/head
2025-12-04T08:53:09.5487003Z  * [new branch]                gh/pianpwk/30/orig      -> origin/gh/pianpwk/30/orig
2025-12-04T08:53:09.5487072Z  * [new branch]                gh/pianpwk/31/base      -> origin/gh/pianpwk/31/base
2025-12-04T08:53:09.5487149Z  * [new branch]                gh/pianpwk/31/head      -> origin/gh/pianpwk/31/head
2025-12-04T08:53:09.5487220Z  * [new branch]                gh/pianpwk/31/orig      -> origin/gh/pianpwk/31/orig
2025-12-04T08:53:09.5487289Z  * [new branch]                gh/pianpwk/32/base      -> origin/gh/pianpwk/32/base
2025-12-04T08:53:09.5487363Z  * [new branch]                gh/pianpwk/32/head      -> origin/gh/pianpwk/32/head
2025-12-04T08:53:09.5487434Z  * [new branch]                gh/pianpwk/32/orig      -> origin/gh/pianpwk/32/orig
2025-12-04T08:53:09.5487538Z  * [new branch]                gh/pianpwk/33/base      -> origin/gh/pianpwk/33/base
2025-12-04T08:53:09.5487635Z  * [new branch]                gh/pianpwk/33/head      -> origin/gh/pianpwk/33/head
2025-12-04T08:53:09.5487706Z  * [new branch]                gh/pianpwk/33/orig      -> origin/gh/pianpwk/33/orig
2025-12-04T08:53:09.5490646Z  * [new branch]                gh/pianpwk/34/base      -> origin/gh/pianpwk/34/base
2025-12-04T08:53:09.5490729Z  * [new branch]                gh/pianpwk/34/head      -> origin/gh/pianpwk/34/head
2025-12-04T08:53:09.5490799Z  * [new branch]                gh/pianpwk/34/orig      -> origin/gh/pianpwk/34/orig
2025-12-04T08:53:09.5490869Z  * [new branch]                gh/pianpwk/35/base      -> origin/gh/pianpwk/35/base
2025-12-04T08:53:09.5490937Z  * [new branch]                gh/pianpwk/35/head      -> origin/gh/pianpwk/35/head
2025-12-04T08:53:09.5491007Z  * [new branch]                gh/pianpwk/35/orig      -> origin/gh/pianpwk/35/orig
2025-12-04T08:53:09.5491082Z  * [new branch]                gh/rec/141/base         -> origin/gh/rec/141/base
2025-12-04T08:53:09.5491154Z  * [new branch]                gh/rec/141/head         -> origin/gh/rec/141/head
2025-12-04T08:53:09.5491218Z  * [new branch]                gh/rec/153/base         -> origin/gh/rec/153/base
2025-12-04T08:53:09.5491281Z  * [new branch]                gh/rec/153/head         -> origin/gh/rec/153/head
2025-12-04T08:53:09.5491345Z  * [new branch]                gh/rec/153/orig         -> origin/gh/rec/153/orig
2025-12-04T08:53:09.5491407Z  * [new branch]                gh/rec/154/base         -> origin/gh/rec/154/base
2025-12-04T08:53:09.5491470Z  * [new branch]                gh/rec/154/head         -> origin/gh/rec/154/head
2025-12-04T08:53:09.5491533Z  * [new branch]                gh/rec/154/orig         -> origin/gh/rec/154/orig
2025-12-04T08:53:09.5491595Z  * [new branch]                gh/rec/164/base         -> origin/gh/rec/164/base
2025-12-04T08:53:09.5491661Z  * [new branch]                gh/rec/164/head         -> origin/gh/rec/164/head
2025-12-04T08:53:09.5491724Z  * [new branch]                gh/rec/164/orig         -> origin/gh/rec/164/orig
2025-12-04T08:53:09.5491792Z  * [new branch]                gh/rec/166/base         -> origin/gh/rec/166/base
2025-12-04T08:53:09.5491854Z  * [new branch]                gh/rec/166/head         -> origin/gh/rec/166/head
2025-12-04T08:53:09.5491920Z  * [new branch]                gh/rec/166/orig         -> origin/gh/rec/166/orig
2025-12-04T08:53:09.5491982Z  * [new branch]                gh/rec/167/base         -> origin/gh/rec/167/base
2025-12-04T08:53:09.5492045Z  * [new branch]                gh/rec/167/head         -> origin/gh/rec/167/head
2025-12-04T08:53:09.5492108Z  * [new branch]                gh/rec/167/orig         -> origin/gh/rec/167/orig
2025-12-04T08:53:09.5492171Z  * [new branch]                gh/rec/168/base         -> origin/gh/rec/168/base
2025-12-04T08:53:09.5492232Z  * [new branch]                gh/rec/168/head         -> origin/gh/rec/168/head
2025-12-04T08:53:09.5492297Z  * [new branch]                gh/rec/168/orig         -> origin/gh/rec/168/orig
2025-12-04T08:53:09.5492361Z  * [new branch]                gh/rec/169/base         -> origin/gh/rec/169/base
2025-12-04T08:53:09.5492424Z  * [new branch]                gh/rec/169/head         -> origin/gh/rec/169/head
2025-12-04T08:53:09.5492487Z  * [new branch]                gh/rec/169/orig         -> origin/gh/rec/169/orig
2025-12-04T08:53:09.5492549Z  * [new branch]                gh/rec/170/base         -> origin/gh/rec/170/base
2025-12-04T08:53:09.5492611Z  * [new branch]                gh/rec/170/head         -> origin/gh/rec/170/head
2025-12-04T08:53:09.5492674Z  * [new branch]                gh/rec/170/orig         -> origin/gh/rec/170/orig
2025-12-04T08:53:09.5492736Z  * [new branch]                gh/rec/171/base         -> origin/gh/rec/171/base
2025-12-04T08:53:09.5492798Z  * [new branch]                gh/rec/171/head         -> origin/gh/rec/171/head
2025-12-04T08:53:09.5492920Z  * [new branch]                gh/rec/171/orig         -> origin/gh/rec/171/orig
2025-12-04T08:53:09.5493022Z  * [new branch]                gh/rec/172/base         -> origin/gh/rec/172/base
2025-12-04T08:53:09.5493086Z  * [new branch]                gh/rec/172/head         -> origin/gh/rec/172/head
2025-12-04T08:53:09.5493148Z  * [new branch]                gh/rec/172/orig         -> origin/gh/rec/172/orig
2025-12-04T08:53:09.5493212Z  * [new branch]                gh/rec/173/base         -> origin/gh/rec/173/base
2025-12-04T08:53:09.5493275Z  * [new branch]                gh/rec/173/head         -> origin/gh/rec/173/head
2025-12-04T08:53:09.5493338Z  * [new branch]                gh/rec/173/orig         -> origin/gh/rec/173/orig
2025-12-04T08:53:09.5493400Z  * [new branch]                gh/rec/174/base         -> origin/gh/rec/174/base
2025-12-04T08:53:09.5493464Z  * [new branch]                gh/rec/174/head         -> origin/gh/rec/174/head
2025-12-04T08:53:09.5493531Z  * [new branch]                gh/rec/174/orig         -> origin/gh/rec/174/orig
2025-12-04T08:53:09.5493594Z  * [new branch]                gh/rec/175/base         -> origin/gh/rec/175/base
2025-12-04T08:53:09.5493660Z  * [new branch]                gh/rec/175/head         -> origin/gh/rec/175/head
2025-12-04T08:53:09.5493723Z  * [new branch]                gh/rec/175/orig         -> origin/gh/rec/175/orig
2025-12-04T08:53:09.5493785Z  * [new branch]                gh/rec/176/base         -> origin/gh/rec/176/base
2025-12-04T08:53:09.5493848Z  * [new branch]                gh/rec/176/head         -> origin/gh/rec/176/head
2025-12-04T08:53:09.5493910Z  * [new branch]                gh/rec/176/orig         -> origin/gh/rec/176/orig
2025-12-04T08:53:09.5493972Z  * [new branch]                gh/rec/177/base         -> origin/gh/rec/177/base
2025-12-04T08:53:09.5494035Z  * [new branch]                gh/rec/177/head         -> origin/gh/rec/177/head
2025-12-04T08:53:09.5494099Z  * [new branch]                gh/rec/177/orig         -> origin/gh/rec/177/orig
2025-12-04T08:53:09.5494190Z  * [new branch]                gh/robert-hardwick/3/base -> origin/gh/robert-hardwick/3/base
2025-12-04T08:53:09.5494278Z  * [new branch]                gh/robert-hardwick/3/head -> origin/gh/robert-hardwick/3/head
2025-12-04T08:53:09.5494360Z  * [new branch]                gh/robert-hardwick/3/orig -> origin/gh/robert-hardwick/3/orig
2025-12-04T08:53:09.5494440Z  * [new branch]                gh/robert-hardwick/4/base -> origin/gh/robert-hardwick/4/base
2025-12-04T08:53:09.5494521Z  * [new branch]                gh/robert-hardwick/4/head -> origin/gh/robert-hardwick/4/head
2025-12-04T08:53:09.5494601Z  * [new branch]                gh/robert-hardwick/4/orig -> origin/gh/robert-hardwick/4/orig
2025-12-04T08:53:09.5494683Z  * [new branch]                gh/robert-hardwick/5/base -> origin/gh/robert-hardwick/5/base
2025-12-04T08:53:09.5494762Z  * [new branch]                gh/robert-hardwick/5/head -> origin/gh/robert-hardwick/5/head
2025-12-04T08:53:09.5494845Z  * [new branch]                gh/robert-hardwick/5/orig -> origin/gh/robert-hardwick/5/orig
2025-12-04T08:53:09.5494930Z  * [new branch]                gh/robert-hardwick/6/base -> origin/gh/robert-hardwick/6/base
2025-12-04T08:53:09.5495011Z  * [new branch]                gh/robert-hardwick/6/head -> origin/gh/robert-hardwick/6/head
2025-12-04T08:53:09.5495092Z  * [new branch]                gh/robert-hardwick/6/orig -> origin/gh/robert-hardwick/6/orig
2025-12-04T08:53:09.5495173Z  * [new branch]                gh/robert-hardwick/7/base -> origin/gh/robert-hardwick/7/base
2025-12-04T08:53:09.5495255Z  * [new branch]                gh/robert-hardwick/7/head -> origin/gh/robert-hardwick/7/head
2025-12-04T08:53:09.5495335Z  * [new branch]                gh/robert-hardwick/7/orig -> origin/gh/robert-hardwick/7/orig
2025-12-04T08:53:09.5495416Z  * [new branch]                gh/robert-hardwick/8/base -> origin/gh/robert-hardwick/8/base
2025-12-04T08:53:09.5495523Z  * [new branch]                gh/robert-hardwick/8/head -> origin/gh/robert-hardwick/8/head
2025-12-04T08:53:09.5495604Z  * [new branch]                gh/robert-hardwick/8/orig -> origin/gh/robert-hardwick/8/orig
2025-12-04T08:53:09.5495714Z  * [new branch]                gh/robert-hardwick/9/base -> origin/gh/robert-hardwick/9/base
2025-12-04T08:53:09.5495797Z  * [new branch]                gh/robert-hardwick/9/head -> origin/gh/robert-hardwick/9/head
2025-12-04T08:53:09.5495878Z  * [new branch]                gh/robert-hardwick/9/orig -> origin/gh/robert-hardwick/9/orig
2025-12-04T08:53:09.5495948Z  * [new branch]                gh/rtimpe/1/base        -> origin/gh/rtimpe/1/base
2025-12-04T08:53:09.5496016Z  * [new branch]                gh/rtimpe/1/head        -> origin/gh/rtimpe/1/head
2025-12-04T08:53:09.5496083Z  * [new branch]                gh/rtimpe/2/base        -> origin/gh/rtimpe/2/base
2025-12-04T08:53:09.5496148Z  * [new branch]                gh/rtimpe/2/head        -> origin/gh/rtimpe/2/head
2025-12-04T08:53:09.5496219Z  * [new branch]                gh/rtimpe/22/base       -> origin/gh/rtimpe/22/base
2025-12-04T08:53:09.5496286Z  * [new branch]                gh/rtimpe/22/head       -> origin/gh/rtimpe/22/head
2025-12-04T08:53:09.5496355Z  * [new branch]                gh/rtimpe/22/orig       -> origin/gh/rtimpe/22/orig
2025-12-04T08:53:09.5496423Z  * [new branch]                gh/rtimpe/23/base       -> origin/gh/rtimpe/23/base
2025-12-04T08:53:09.5496489Z  * [new branch]                gh/rtimpe/23/head       -> origin/gh/rtimpe/23/head
2025-12-04T08:53:09.5496555Z  * [new branch]                gh/rtimpe/23/orig       -> origin/gh/rtimpe/23/orig
2025-12-04T08:53:09.5496622Z  * [new branch]                gh/rtimpe/24/base       -> origin/gh/rtimpe/24/base
2025-12-04T08:53:09.5496688Z  * [new branch]                gh/rtimpe/24/head       -> origin/gh/rtimpe/24/head
2025-12-04T08:53:09.5496753Z  * [new branch]                gh/rtimpe/24/orig       -> origin/gh/rtimpe/24/orig
2025-12-04T08:53:09.5496821Z  * [new branch]                gh/rtimpe/25/base       -> origin/gh/rtimpe/25/base
2025-12-04T08:53:09.5496887Z  * [new branch]                gh/rtimpe/25/head       -> origin/gh/rtimpe/25/head
2025-12-04T08:53:09.5496956Z  * [new branch]                gh/rtimpe/25/orig       -> origin/gh/rtimpe/25/orig
2025-12-04T08:53:09.5497022Z  * [new branch]                gh/rtimpe/26/base       -> origin/gh/rtimpe/26/base
2025-12-04T08:53:09.5497087Z  * [new branch]                gh/rtimpe/26/head       -> origin/gh/rtimpe/26/head
2025-12-04T08:53:09.5497154Z  * [new branch]                gh/rtimpe/26/orig       -> origin/gh/rtimpe/26/orig
2025-12-04T08:53:09.5497219Z  * [new branch]                gh/rtimpe/27/base       -> origin/gh/rtimpe/27/base
2025-12-04T08:53:09.5497285Z  * [new branch]                gh/rtimpe/27/head       -> origin/gh/rtimpe/27/head
2025-12-04T08:53:09.5497351Z  * [new branch]                gh/rtimpe/27/orig       -> origin/gh/rtimpe/27/orig
2025-12-04T08:53:09.5497418Z  * [new branch]                gh/rtimpe/28/base       -> origin/gh/rtimpe/28/base
2025-12-04T08:53:09.5497484Z  * [new branch]                gh/rtimpe/28/head       -> origin/gh/rtimpe/28/head
2025-12-04T08:53:09.5497552Z  * [new branch]                gh/rtimpe/28/orig       -> origin/gh/rtimpe/28/orig
2025-12-04T08:53:09.5497618Z  * [new branch]                gh/rtimpe/29/base       -> origin/gh/rtimpe/29/base
2025-12-04T08:53:09.5497684Z  * [new branch]                gh/rtimpe/29/head       -> origin/gh/rtimpe/29/head
2025-12-04T08:53:09.5497750Z  * [new branch]                gh/rtimpe/29/orig       -> origin/gh/rtimpe/29/orig
2025-12-04T08:53:09.5497816Z  * [new branch]                gh/rtimpe/3/base        -> origin/gh/rtimpe/3/base
2025-12-04T08:53:09.5497882Z  * [new branch]                gh/rtimpe/3/head        -> origin/gh/rtimpe/3/head
2025-12-04T08:53:09.5497950Z  * [new branch]                gh/rtimpe/30/base       -> origin/gh/rtimpe/30/base
2025-12-04T08:53:09.5498015Z  * [new branch]                gh/rtimpe/30/head       -> origin/gh/rtimpe/30/head
2025-12-04T08:53:09.5498108Z  * [new branch]                gh/rtimpe/30/orig       -> origin/gh/rtimpe/30/orig
2025-12-04T08:53:09.5498206Z  * [new branch]                gh/rtimpe/31/base       -> origin/gh/rtimpe/31/base
2025-12-04T08:53:09.5498273Z  * [new branch]                gh/rtimpe/31/head       -> origin/gh/rtimpe/31/head
2025-12-04T08:53:09.5498338Z  * [new branch]                gh/rtimpe/31/orig       -> origin/gh/rtimpe/31/orig
2025-12-04T08:53:09.5498405Z  * [new branch]                gh/rtimpe/32/base       -> origin/gh/rtimpe/32/base
2025-12-04T08:53:09.5498471Z  * [new branch]                gh/rtimpe/32/head       -> origin/gh/rtimpe/32/head
2025-12-04T08:53:09.5498538Z  * [new branch]                gh/rtimpe/32/orig       -> origin/gh/rtimpe/32/orig
2025-12-04T08:53:09.5498604Z  * [new branch]                gh/rtimpe/33/base       -> origin/gh/rtimpe/33/base
2025-12-04T08:53:09.5498670Z  * [new branch]                gh/rtimpe/33/head       -> origin/gh/rtimpe/33/head
2025-12-04T08:53:09.5498739Z  * [new branch]                gh/rtimpe/33/orig       -> origin/gh/rtimpe/33/orig
2025-12-04T08:53:09.5498806Z  * [new branch]                gh/rtimpe/34/base       -> origin/gh/rtimpe/34/base
2025-12-04T08:53:09.5498872Z  * [new branch]                gh/rtimpe/34/head       -> origin/gh/rtimpe/34/head
2025-12-04T08:53:09.5498939Z  * [new branch]                gh/rtimpe/34/orig       -> origin/gh/rtimpe/34/orig
2025-12-04T08:53:09.5499004Z  * [new branch]                gh/rtimpe/35/base       -> origin/gh/rtimpe/35/base
2025-12-04T08:53:09.5499070Z  * [new branch]                gh/rtimpe/35/head       -> origin/gh/rtimpe/35/head
2025-12-04T08:53:09.5499137Z  * [new branch]                gh/rtimpe/35/orig       -> origin/gh/rtimpe/35/orig
2025-12-04T08:53:09.5499203Z  * [new branch]                gh/rtimpe/4/base        -> origin/gh/rtimpe/4/base
2025-12-04T08:53:09.5499269Z  * [new branch]                gh/rtimpe/4/head        -> origin/gh/rtimpe/4/head
2025-12-04T08:53:09.5499353Z  * [new branch]                gh/ruisizhang123/1/base -> origin/gh/ruisizhang123/1/base
2025-12-04T08:53:09.5499431Z  * [new branch]                gh/ruisizhang123/1/head -> origin/gh/ruisizhang123/1/head
2025-12-04T08:53:09.5499508Z  * [new branch]                gh/ruisizhang123/1/orig -> origin/gh/ruisizhang123/1/orig
2025-12-04T08:53:09.5499585Z  * [new branch]                gh/ruisizhang123/4/base -> origin/gh/ruisizhang123/4/base
2025-12-04T08:53:09.5499660Z  * [new branch]                gh/ruisizhang123/4/head -> origin/gh/ruisizhang123/4/head
2025-12-04T08:53:09.5499735Z  * [new branch]                gh/ruisizhang123/4/orig -> origin/gh/ruisizhang123/4/orig
2025-12-04T08:53:09.5499811Z  * [new branch]                gh/ruisizhang123/5/base -> origin/gh/ruisizhang123/5/base
2025-12-04T08:53:09.5499887Z  * [new branch]                gh/ruisizhang123/5/head -> origin/gh/ruisizhang123/5/head
2025-12-04T08:53:09.5499961Z  * [new branch]                gh/ruisizhang123/5/orig -> origin/gh/ruisizhang123/5/orig
2025-12-04T08:53:09.5500038Z  * [new branch]                gh/ruisizhang123/6/base -> origin/gh/ruisizhang123/6/base
2025-12-04T08:53:09.5500115Z  * [new branch]                gh/ruisizhang123/6/head -> origin/gh/ruisizhang123/6/head
2025-12-04T08:53:09.5500190Z  * [new branch]                gh/ruisizhang123/6/orig -> origin/gh/ruisizhang123/6/orig
2025-12-04T08:53:09.5500265Z  * [new branch]                gh/ruisizhang123/7/base -> origin/gh/ruisizhang123/7/base
2025-12-04T08:53:09.5500339Z  * [new branch]                gh/ruisizhang123/7/head -> origin/gh/ruisizhang123/7/head
2025-12-04T08:53:09.5500451Z  * [new branch]                gh/ruisizhang123/7/orig -> origin/gh/ruisizhang123/7/orig
2025-12-04T08:53:09.5500526Z  * [new branch]                gh/ruisizhang123/8/base -> origin/gh/ruisizhang123/8/base
2025-12-04T08:53:09.5500601Z  * [new branch]                gh/ruisizhang123/8/head -> origin/gh/ruisizhang123/8/head
2025-12-04T08:53:09.5500713Z  * [new branch]                gh/ruisizhang123/8/orig -> origin/gh/ruisizhang123/8/orig
2025-12-04T08:53:09.5500789Z  * [new branch]                gh/ruisizhang123/9/base -> origin/gh/ruisizhang123/9/base
2025-12-04T08:53:09.5500905Z  * [new branch]                gh/ruisizhang123/9/head -> origin/gh/ruisizhang123/9/head
2025-12-04T08:53:09.5500981Z  * [new branch]                gh/ruisizhang123/9/orig -> origin/gh/ruisizhang123/9/orig
2025-12-04T08:53:09.5501057Z  * [new branch]                gh/seemethere/52/base   -> origin/gh/seemethere/52/base
2025-12-04T08:53:09.5501131Z  * [new branch]                gh/seemethere/52/head   -> origin/gh/seemethere/52/head
2025-12-04T08:53:09.5501206Z  * [new branch]                gh/seemethere/52/orig   -> origin/gh/seemethere/52/orig
2025-12-04T08:53:09.5501279Z  * [new branch]                gh/seemethere/53/base   -> origin/gh/seemethere/53/base
2025-12-04T08:53:09.5501352Z  * [new branch]                gh/seemethere/53/head   -> origin/gh/seemethere/53/head
2025-12-04T08:53:09.5501427Z  * [new branch]                gh/seemethere/53/orig   -> origin/gh/seemethere/53/orig
2025-12-04T08:53:09.5501500Z  * [new branch]                gh/seemethere/54/base   -> origin/gh/seemethere/54/base
2025-12-04T08:53:09.5501572Z  * [new branch]                gh/seemethere/54/head   -> origin/gh/seemethere/54/head
2025-12-04T08:53:09.5501646Z  * [new branch]                gh/seemethere/54/orig   -> origin/gh/seemethere/54/orig
2025-12-04T08:53:09.5501719Z  * [new branch]                gh/seemethere/55/base   -> origin/gh/seemethere/55/base
2025-12-04T08:53:09.5501791Z  * [new branch]                gh/seemethere/55/head   -> origin/gh/seemethere/55/head
2025-12-04T08:53:09.5501864Z  * [new branch]                gh/seemethere/55/orig   -> origin/gh/seemethere/55/orig
2025-12-04T08:53:09.5501936Z  * [new branch]                gh/seemethere/59/base   -> origin/gh/seemethere/59/base
2025-12-04T08:53:09.5502009Z  * [new branch]                gh/seemethere/59/head   -> origin/gh/seemethere/59/head
2025-12-04T08:53:09.5502083Z  * [new branch]                gh/seemethere/59/orig   -> origin/gh/seemethere/59/orig
2025-12-04T08:53:09.5502160Z  * [new branch]                gh/seemethere/62/base   -> origin/gh/seemethere/62/base
2025-12-04T08:53:09.5502234Z  * [new branch]                gh/seemethere/62/head   -> origin/gh/seemethere/62/head
2025-12-04T08:53:09.5502305Z  * [new branch]                gh/seemethere/62/orig   -> origin/gh/seemethere/62/orig
2025-12-04T08:53:09.5502378Z  * [new branch]                gh/seemethere/63/base   -> origin/gh/seemethere/63/base
2025-12-04T08:53:09.5502451Z  * [new branch]                gh/seemethere/63/head   -> origin/gh/seemethere/63/head
2025-12-04T08:53:09.5502523Z  * [new branch]                gh/seemethere/63/orig   -> origin/gh/seemethere/63/orig
2025-12-04T08:53:09.5502596Z  * [new branch]                gh/seemethere/71/base   -> origin/gh/seemethere/71/base
2025-12-04T08:53:09.5502670Z  * [new branch]                gh/seemethere/71/head   -> origin/gh/seemethere/71/head
2025-12-04T08:53:09.5502744Z  * [new branch]                gh/seemethere/71/orig   -> origin/gh/seemethere/71/orig
2025-12-04T08:53:09.5502818Z  * [new branch]                gh/seemethere/72/base   -> origin/gh/seemethere/72/base
2025-12-04T08:53:09.5502893Z  * [new branch]                gh/seemethere/72/head   -> origin/gh/seemethere/72/head
2025-12-04T08:53:09.5502964Z  * [new branch]                gh/seemethere/72/orig   -> origin/gh/seemethere/72/orig
2025-12-04T08:53:09.5503036Z  * [new branch]                gh/seemethere/73/base   -> origin/gh/seemethere/73/base
2025-12-04T08:53:09.5503109Z  * [new branch]                gh/seemethere/73/head   -> origin/gh/seemethere/73/head
2025-12-04T08:53:09.5503180Z  * [new branch]                gh/seemethere/73/orig   -> origin/gh/seemethere/73/orig
2025-12-04T08:53:09.5503252Z  * [new branch]                gh/seemethere/74/base   -> origin/gh/seemethere/74/base
2025-12-04T08:53:09.5503352Z  * [new branch]                gh/seemethere/74/head   -> origin/gh/seemethere/74/head
2025-12-04T08:53:09.5503425Z  * [new branch]                gh/seemethere/74/orig   -> origin/gh/seemethere/74/orig
2025-12-04T08:53:09.5503522Z  * [new branch]                gh/seemethere/75/base   -> origin/gh/seemethere/75/base
2025-12-04T08:53:09.5503595Z  * [new branch]                gh/seemethere/75/head   -> origin/gh/seemethere/75/head
2025-12-04T08:53:09.5503668Z  * [new branch]                gh/seemethere/75/orig   -> origin/gh/seemethere/75/orig
2025-12-04T08:53:09.5503741Z  * [new branch]                gh/seemethere/76/base   -> origin/gh/seemethere/76/base
2025-12-04T08:53:09.5503814Z  * [new branch]                gh/seemethere/76/head   -> origin/gh/seemethere/76/head
2025-12-04T08:53:09.5503886Z  * [new branch]                gh/seemethere/76/orig   -> origin/gh/seemethere/76/orig
2025-12-04T08:53:09.5503963Z  * [new branch]                gh/shunting314/145/base -> origin/gh/shunting314/145/base
2025-12-04T08:53:09.5504041Z  * [new branch]                gh/shunting314/145/head -> origin/gh/shunting314/145/head
2025-12-04T08:53:09.5504116Z  * [new branch]                gh/shunting314/145/orig -> origin/gh/shunting314/145/orig
2025-12-04T08:53:09.5504192Z  * [new branch]                gh/shunting314/176/base -> origin/gh/shunting314/176/base
2025-12-04T08:53:09.5504266Z  * [new branch]                gh/shunting314/176/head -> origin/gh/shunting314/176/head
2025-12-04T08:53:09.5504340Z  * [new branch]                gh/shunting314/176/orig -> origin/gh/shunting314/176/orig
2025-12-04T08:53:09.5504415Z  * [new branch]                gh/shunting314/249/base -> origin/gh/shunting314/249/base
2025-12-04T08:53:09.5504488Z  * [new branch]                gh/shunting314/249/head -> origin/gh/shunting314/249/head
2025-12-04T08:53:09.5504561Z  * [new branch]                gh/shunting314/249/orig -> origin/gh/shunting314/249/orig
2025-12-04T08:53:09.5504635Z  * [new branch]                gh/shunting314/253/base -> origin/gh/shunting314/253/base
2025-12-04T08:53:09.5504710Z  * [new branch]                gh/shunting314/253/head -> origin/gh/shunting314/253/head
2025-12-04T08:53:09.5504784Z  * [new branch]                gh/shunting314/253/orig -> origin/gh/shunting314/253/orig
2025-12-04T08:53:09.5504859Z  * [new branch]                gh/shunting314/256/base -> origin/gh/shunting314/256/base
2025-12-04T08:53:09.5504932Z  * [new branch]                gh/shunting314/256/head -> origin/gh/shunting314/256/head
2025-12-04T08:53:09.5505007Z  * [new branch]                gh/shunting314/256/orig -> origin/gh/shunting314/256/orig
2025-12-04T08:53:09.5505081Z  * [new branch]                gh/shunting314/257/base -> origin/gh/shunting314/257/base
2025-12-04T08:53:09.5505155Z  * [new branch]                gh/shunting314/257/head -> origin/gh/shunting314/257/head
2025-12-04T08:53:09.5505229Z  * [new branch]                gh/shunting314/257/orig -> origin/gh/shunting314/257/orig
2025-12-04T08:53:09.5505303Z  * [new branch]                gh/shunting314/258/base -> origin/gh/shunting314/258/base
2025-12-04T08:53:09.5505377Z  * [new branch]                gh/shunting314/258/head -> origin/gh/shunting314/258/head
2025-12-04T08:53:09.5505453Z  * [new branch]                gh/shunting314/258/orig -> origin/gh/shunting314/258/orig
2025-12-04T08:53:09.5505527Z  * [new branch]                gh/shunting314/259/base -> origin/gh/shunting314/259/base
2025-12-04T08:53:09.5505601Z  * [new branch]                gh/shunting314/259/head -> origin/gh/shunting314/259/head
2025-12-04T08:53:09.5505675Z  * [new branch]                gh/shunting314/259/orig -> origin/gh/shunting314/259/orig
2025-12-04T08:53:09.5505749Z  * [new branch]                gh/shunting314/260/base -> origin/gh/shunting314/260/base
2025-12-04T08:53:09.5505822Z  * [new branch]                gh/shunting314/260/head -> origin/gh/shunting314/260/head
2025-12-04T08:53:09.5505897Z  * [new branch]                gh/shunting314/260/orig -> origin/gh/shunting314/260/orig
2025-12-04T08:53:09.5506004Z  * [new branch]                gh/shunting314/261/base -> origin/gh/shunting314/261/base
2025-12-04T08:53:09.5506078Z  * [new branch]                gh/shunting314/261/head -> origin/gh/shunting314/261/head
2025-12-04T08:53:09.5506179Z  * [new branch]                gh/shunting314/261/orig -> origin/gh/shunting314/261/orig
2025-12-04T08:53:09.5506252Z  * [new branch]                gh/shunting314/262/base -> origin/gh/shunting314/262/base
2025-12-04T08:53:09.5506325Z  * [new branch]                gh/shunting314/262/head -> origin/gh/shunting314/262/head
2025-12-04T08:53:09.5506400Z  * [new branch]                gh/shunting314/262/orig -> origin/gh/shunting314/262/orig
2025-12-04T08:53:09.5506473Z  * [new branch]                gh/shunting314/263/base -> origin/gh/shunting314/263/base
2025-12-04T08:53:09.5506547Z  * [new branch]                gh/shunting314/263/head -> origin/gh/shunting314/263/head
2025-12-04T08:53:09.5506622Z  * [new branch]                gh/shunting314/263/orig -> origin/gh/shunting314/263/orig
2025-12-04T08:53:09.5506697Z  * [new branch]                gh/shunting314/264/base -> origin/gh/shunting314/264/base
2025-12-04T08:53:09.5506773Z  * [new branch]                gh/shunting314/264/head -> origin/gh/shunting314/264/head
2025-12-04T08:53:09.5506846Z  * [new branch]                gh/shunting314/264/orig -> origin/gh/shunting314/264/orig
2025-12-04T08:53:09.5506920Z  * [new branch]                gh/shunting314/265/base -> origin/gh/shunting314/265/base
2025-12-04T08:53:09.5506994Z  * [new branch]                gh/shunting314/265/head -> origin/gh/shunting314/265/head
2025-12-04T08:53:09.5507068Z  * [new branch]                gh/shunting314/265/orig -> origin/gh/shunting314/265/orig
2025-12-04T08:53:09.5507141Z  * [new branch]                gh/shunting314/266/base -> origin/gh/shunting314/266/base
2025-12-04T08:53:09.5507215Z  * [new branch]                gh/shunting314/266/head -> origin/gh/shunting314/266/head
2025-12-04T08:53:09.5507289Z  * [new branch]                gh/shunting314/266/orig -> origin/gh/shunting314/266/orig
2025-12-04T08:53:09.5507363Z  * [new branch]                gh/shunting314/267/base -> origin/gh/shunting314/267/base
2025-12-04T08:53:09.5507438Z  * [new branch]                gh/shunting314/267/head -> origin/gh/shunting314/267/head
2025-12-04T08:53:09.5507512Z  * [new branch]                gh/shunting314/267/orig -> origin/gh/shunting314/267/orig
2025-12-04T08:53:09.5507585Z  * [new branch]                gh/shunting314/268/base -> origin/gh/shunting314/268/base
2025-12-04T08:53:09.5507660Z  * [new branch]                gh/shunting314/268/head -> origin/gh/shunting314/268/head
2025-12-04T08:53:09.5507733Z  * [new branch]                gh/shunting314/268/orig -> origin/gh/shunting314/268/orig
2025-12-04T08:53:09.5507807Z  * [new branch]                gh/shunting314/269/base -> origin/gh/shunting314/269/base
2025-12-04T08:53:09.5507881Z  * [new branch]                gh/shunting314/269/head -> origin/gh/shunting314/269/head
2025-12-04T08:53:09.5507958Z  * [new branch]                gh/shunting314/269/orig -> origin/gh/shunting314/269/orig
2025-12-04T08:53:09.5508031Z  * [new branch]                gh/silverguo/1/base     -> origin/gh/silverguo/1/base
2025-12-04T08:53:09.5508106Z  * [new branch]                gh/silverguo/1/head     -> origin/gh/silverguo/1/head
2025-12-04T08:53:09.5508178Z  * [new branch]                gh/silverguo/2/base     -> origin/gh/silverguo/2/base
2025-12-04T08:53:09.5508249Z  * [new branch]                gh/silverguo/2/head     -> origin/gh/silverguo/2/head
2025-12-04T08:53:09.5508318Z  * [new branch]                gh/silverguo/3/base     -> origin/gh/silverguo/3/base
2025-12-04T08:53:09.5508387Z  * [new branch]                gh/silverguo/3/head     -> origin/gh/silverguo/3/head
2025-12-04T08:53:09.5508456Z  * [new branch]                gh/silverguo/4/base     -> origin/gh/silverguo/4/base
2025-12-04T08:53:09.5508525Z  * [new branch]                gh/silverguo/4/head     -> origin/gh/silverguo/4/head
2025-12-04T08:53:09.5508625Z  * [new branch]                gh/slayton58/39/base    -> origin/gh/slayton58/39/base
2025-12-04T08:53:09.5508698Z  * [new branch]                gh/slayton58/39/head    -> origin/gh/slayton58/39/head
2025-12-04T08:53:09.5508804Z  * [new branch]                gh/slayton58/39/orig    -> origin/gh/slayton58/39/orig
2025-12-04T08:53:09.5508874Z  * [new branch]                gh/slayton58/42/base    -> origin/gh/slayton58/42/base
2025-12-04T08:53:09.5508945Z  * [new branch]                gh/slayton58/42/head    -> origin/gh/slayton58/42/head
2025-12-04T08:53:09.5509014Z  * [new branch]                gh/slayton58/42/orig    -> origin/gh/slayton58/42/orig
2025-12-04T08:53:09.5509083Z  * [new branch]                gh/slayton58/43/base    -> origin/gh/slayton58/43/base
2025-12-04T08:53:09.5509153Z  * [new branch]                gh/slayton58/43/head    -> origin/gh/slayton58/43/head
2025-12-04T08:53:09.5509223Z  * [new branch]                gh/slayton58/43/orig    -> origin/gh/slayton58/43/orig
2025-12-04T08:53:09.5509296Z  * [new branch]                gh/slayton58/44/base    -> origin/gh/slayton58/44/base
2025-12-04T08:53:09.5509369Z  * [new branch]                gh/slayton58/44/head    -> origin/gh/slayton58/44/head
2025-12-04T08:53:09.5509439Z  * [new branch]                gh/slayton58/44/orig    -> origin/gh/slayton58/44/orig
2025-12-04T08:53:09.5509509Z  * [new branch]                gh/slayton58/45/base    -> origin/gh/slayton58/45/base
2025-12-04T08:53:09.5509580Z  * [new branch]                gh/slayton58/45/head    -> origin/gh/slayton58/45/head
2025-12-04T08:53:09.5509649Z  * [new branch]                gh/slayton58/45/orig    -> origin/gh/slayton58/45/orig
2025-12-04T08:53:09.5509719Z  * [new branch]                gh/slayton58/46/base    -> origin/gh/slayton58/46/base
2025-12-04T08:53:09.5509788Z  * [new branch]                gh/slayton58/46/head    -> origin/gh/slayton58/46/head
2025-12-04T08:53:09.5509858Z  * [new branch]                gh/slayton58/46/orig    -> origin/gh/slayton58/46/orig
2025-12-04T08:53:09.5509933Z  * [new branch]                gh/slayton58/6/base     -> origin/gh/slayton58/6/base
2025-12-04T08:53:09.5510002Z  * [new branch]                gh/slayton58/6/head     -> origin/gh/slayton58/6/head
2025-12-04T08:53:09.5510072Z  * [new branch]                gh/slayton58/7/base     -> origin/gh/slayton58/7/base
2025-12-04T08:53:09.5510142Z  * [new branch]                gh/slayton58/7/head     -> origin/gh/slayton58/7/head
2025-12-04T08:53:09.5510216Z  * [new branch]                gh/soulitzer/269/base   -> origin/gh/soulitzer/269/base
2025-12-04T08:53:09.5510289Z  * [new branch]                gh/soulitzer/269/head   -> origin/gh/soulitzer/269/head
2025-12-04T08:53:09.5510362Z  * [new branch]                gh/soulitzer/269/orig   -> origin/gh/soulitzer/269/orig
2025-12-04T08:53:09.5510466Z  * [new branch]                gh/soulitzer/276/base   -> origin/gh/soulitzer/276/base
2025-12-04T08:53:09.5510539Z  * [new branch]                gh/soulitzer/276/head   -> origin/gh/soulitzer/276/head
2025-12-04T08:53:09.5510614Z  * [new branch]                gh/soulitzer/276/orig   -> origin/gh/soulitzer/276/orig
2025-12-04T08:53:09.5510687Z  * [new branch]                gh/soulitzer/287/base   -> origin/gh/soulitzer/287/base
2025-12-04T08:53:09.5510759Z  * [new branch]                gh/soulitzer/287/head   -> origin/gh/soulitzer/287/head
2025-12-04T08:53:09.5510832Z  * [new branch]                gh/soulitzer/287/orig   -> origin/gh/soulitzer/287/orig
2025-12-04T08:53:09.5510906Z  * [new branch]                gh/soulitzer/296/base   -> origin/gh/soulitzer/296/base
2025-12-04T08:53:09.5510977Z  * [new branch]                gh/soulitzer/296/head   -> origin/gh/soulitzer/296/head
2025-12-04T08:53:09.5511050Z  * [new branch]                gh/soulitzer/296/orig   -> origin/gh/soulitzer/296/orig
2025-12-04T08:53:09.5511121Z  * [new branch]                gh/soulitzer/299/base   -> origin/gh/soulitzer/299/base
2025-12-04T08:53:09.5511193Z  * [new branch]                gh/soulitzer/299/head   -> origin/gh/soulitzer/299/head
2025-12-04T08:53:09.5511317Z  * [new branch]                gh/soulitzer/299/orig   -> origin/gh/soulitzer/299/orig
2025-12-04T08:53:09.5511436Z  * [new branch]                gh/soulitzer/300/base   -> origin/gh/soulitzer/300/base
2025-12-04T08:53:09.5511509Z  * [new branch]                gh/soulitzer/300/head   -> origin/gh/soulitzer/300/head
2025-12-04T08:53:09.5511581Z  * [new branch]                gh/soulitzer/300/orig   -> origin/gh/soulitzer/300/orig
2025-12-04T08:53:09.5511652Z  * [new branch]                gh/soulitzer/301/base   -> origin/gh/soulitzer/301/base
2025-12-04T08:53:09.5511725Z  * [new branch]                gh/soulitzer/301/head   -> origin/gh/soulitzer/301/head
2025-12-04T08:53:09.5511796Z  * [new branch]                gh/soulitzer/301/orig   -> origin/gh/soulitzer/301/orig
2025-12-04T08:53:09.5511868Z  * [new branch]                gh/soulitzer/313/base   -> origin/gh/soulitzer/313/base
2025-12-04T08:53:09.5511940Z  * [new branch]                gh/soulitzer/313/head   -> origin/gh/soulitzer/313/head
2025-12-04T08:53:09.5512015Z  * [new branch]                gh/soulitzer/313/orig   -> origin/gh/soulitzer/313/orig
2025-12-04T08:53:09.5512088Z  * [new branch]                gh/soulitzer/319/base   -> origin/gh/soulitzer/319/base
2025-12-04T08:53:09.5512160Z  * [new branch]                gh/soulitzer/319/head   -> origin/gh/soulitzer/319/head
2025-12-04T08:53:09.5512232Z  * [new branch]                gh/soulitzer/319/orig   -> origin/gh/soulitzer/319/orig
2025-12-04T08:53:09.5512303Z  * [new branch]                gh/soulitzer/320/base   -> origin/gh/soulitzer/320/base
2025-12-04T08:53:09.5512376Z  * [new branch]                gh/soulitzer/320/head   -> origin/gh/soulitzer/320/head
2025-12-04T08:53:09.5512448Z  * [new branch]                gh/soulitzer/320/orig   -> origin/gh/soulitzer/320/orig
2025-12-04T08:53:09.5512519Z  * [new branch]                gh/soulitzer/336/base   -> origin/gh/soulitzer/336/base
2025-12-04T08:53:09.5512593Z  * [new branch]                gh/soulitzer/336/head   -> origin/gh/soulitzer/336/head
2025-12-04T08:53:09.5512665Z  * [new branch]                gh/soulitzer/336/orig   -> origin/gh/soulitzer/336/orig
2025-12-04T08:53:09.5512738Z  * [new branch]                gh/soulitzer/347/base   -> origin/gh/soulitzer/347/base
2025-12-04T08:53:09.5512810Z  * [new branch]                gh/soulitzer/347/head   -> origin/gh/soulitzer/347/head
2025-12-04T08:53:09.5512882Z  * [new branch]                gh/soulitzer/347/orig   -> origin/gh/soulitzer/347/orig
2025-12-04T08:53:09.5512955Z  * [new branch]                gh/soulitzer/349/base   -> origin/gh/soulitzer/349/base
2025-12-04T08:53:09.5513026Z  * [new branch]                gh/soulitzer/349/head   -> origin/gh/soulitzer/349/head
2025-12-04T08:53:09.5513104Z  * [new branch]                gh/soulitzer/349/orig   -> origin/gh/soulitzer/349/orig
2025-12-04T08:53:09.5513176Z  * [new branch]                gh/soulitzer/350/base   -> origin/gh/soulitzer/350/base
2025-12-04T08:53:09.5513251Z  * [new branch]                gh/soulitzer/350/head   -> origin/gh/soulitzer/350/head
2025-12-04T08:53:09.5513323Z  * [new branch]                gh/soulitzer/350/orig   -> origin/gh/soulitzer/350/orig
2025-12-04T08:53:09.5513396Z  * [new branch]                gh/soulitzer/351/base   -> origin/gh/soulitzer/351/base
2025-12-04T08:53:09.5513468Z  * [new branch]                gh/soulitzer/351/head   -> origin/gh/soulitzer/351/head
2025-12-04T08:53:09.5513539Z  * [new branch]                gh/soulitzer/351/orig   -> origin/gh/soulitzer/351/orig
2025-12-04T08:53:09.5513613Z  * [new branch]                gh/soulitzer/353/base   -> origin/gh/soulitzer/353/base
2025-12-04T08:53:09.5513685Z  * [new branch]                gh/soulitzer/353/head   -> origin/gh/soulitzer/353/head
2025-12-04T08:53:09.5513756Z  * [new branch]                gh/soulitzer/353/orig   -> origin/gh/soulitzer/353/orig
2025-12-04T08:53:09.5513829Z  * [new branch]                gh/soulitzer/358/base   -> origin/gh/soulitzer/358/base
2025-12-04T08:53:09.5513929Z  * [new branch]                gh/soulitzer/358/head   -> origin/gh/soulitzer/358/head
2025-12-04T08:53:09.5514000Z  * [new branch]                gh/soulitzer/358/orig   -> origin/gh/soulitzer/358/orig
2025-12-04T08:53:09.5514096Z  * [new branch]                gh/soulitzer/359/base   -> origin/gh/soulitzer/359/base
2025-12-04T08:53:09.5514168Z  * [new branch]                gh/soulitzer/359/head   -> origin/gh/soulitzer/359/head
2025-12-04T08:53:09.5514239Z  * [new branch]                gh/soulitzer/359/orig   -> origin/gh/soulitzer/359/orig
2025-12-04T08:53:09.5514312Z  * [new branch]                gh/soulitzer/374/base   -> origin/gh/soulitzer/374/base
2025-12-04T08:53:09.5514383Z  * [new branch]                gh/soulitzer/374/head   -> origin/gh/soulitzer/374/head
2025-12-04T08:53:09.5514457Z  * [new branch]                gh/soulitzer/374/orig   -> origin/gh/soulitzer/374/orig
2025-12-04T08:53:09.5514528Z  * [new branch]                gh/soulitzer/375/base   -> origin/gh/soulitzer/375/base
2025-12-04T08:53:09.5514602Z  * [new branch]                gh/soulitzer/375/head   -> origin/gh/soulitzer/375/head
2025-12-04T08:53:09.5514676Z  * [new branch]                gh/soulitzer/375/orig   -> origin/gh/soulitzer/375/orig
2025-12-04T08:53:09.5514748Z  * [new branch]                gh/soulitzer/380/base   -> origin/gh/soulitzer/380/base
2025-12-04T08:53:09.5514819Z  * [new branch]                gh/soulitzer/380/head   -> origin/gh/soulitzer/380/head
2025-12-04T08:53:09.5514891Z  * [new branch]                gh/soulitzer/380/orig   -> origin/gh/soulitzer/380/orig
2025-12-04T08:53:09.5514962Z  * [new branch]                gh/soulitzer/385/base   -> origin/gh/soulitzer/385/base
2025-12-04T08:53:09.5515034Z  * [new branch]                gh/soulitzer/385/head   -> origin/gh/soulitzer/385/head
2025-12-04T08:53:09.5515107Z  * [new branch]                gh/soulitzer/385/orig   -> origin/gh/soulitzer/385/orig
2025-12-04T08:53:09.5515179Z  * [new branch]                gh/soulitzer/386/base   -> origin/gh/soulitzer/386/base
2025-12-04T08:53:09.5515252Z  * [new branch]                gh/soulitzer/386/head   -> origin/gh/soulitzer/386/head
2025-12-04T08:53:09.5515325Z  * [new branch]                gh/soulitzer/386/orig   -> origin/gh/soulitzer/386/orig
2025-12-04T08:53:09.5515397Z  * [new branch]                gh/soulitzer/387/base   -> origin/gh/soulitzer/387/base
2025-12-04T08:53:09.5515468Z  * [new branch]                gh/soulitzer/387/head   -> origin/gh/soulitzer/387/head
2025-12-04T08:53:09.5515541Z  * [new branch]                gh/soulitzer/387/orig   -> origin/gh/soulitzer/387/orig
2025-12-04T08:53:09.5515612Z  * [new branch]                gh/soulitzer/388/base   -> origin/gh/soulitzer/388/base
2025-12-04T08:53:09.5515683Z  * [new branch]                gh/soulitzer/388/head   -> origin/gh/soulitzer/388/head
2025-12-04T08:53:09.5515755Z  * [new branch]                gh/soulitzer/388/orig   -> origin/gh/soulitzer/388/orig
2025-12-04T08:53:09.5515826Z  * [new branch]                gh/soulitzer/389/base   -> origin/gh/soulitzer/389/base
2025-12-04T08:53:09.5515899Z  * [new branch]                gh/soulitzer/389/head   -> origin/gh/soulitzer/389/head
2025-12-04T08:53:09.5515972Z  * [new branch]                gh/soulitzer/389/orig   -> origin/gh/soulitzer/389/orig
2025-12-04T08:53:09.5516043Z  * [new branch]                gh/soulitzer/390/base   -> origin/gh/soulitzer/390/base
2025-12-04T08:53:09.5516115Z  * [new branch]                gh/soulitzer/390/head   -> origin/gh/soulitzer/390/head
2025-12-04T08:53:09.5516187Z  * [new branch]                gh/soulitzer/390/orig   -> origin/gh/soulitzer/390/orig
2025-12-04T08:53:09.5516258Z  * [new branch]                gh/soulitzer/391/base   -> origin/gh/soulitzer/391/base
2025-12-04T08:53:09.5516330Z  * [new branch]                gh/soulitzer/391/head   -> origin/gh/soulitzer/391/head
2025-12-04T08:53:09.5516401Z  * [new branch]                gh/soulitzer/391/orig   -> origin/gh/soulitzer/391/orig
2025-12-04T08:53:09.5516501Z  * [new branch]                gh/soulitzer/392/base   -> origin/gh/soulitzer/392/base
2025-12-04T08:53:09.5516574Z  * [new branch]                gh/soulitzer/392/head   -> origin/gh/soulitzer/392/head
2025-12-04T08:53:09.5516675Z  * [new branch]                gh/soulitzer/392/orig   -> origin/gh/soulitzer/392/orig
2025-12-04T08:53:09.5516747Z  * [new branch]                gh/swolchok/728/next    -> origin/gh/swolchok/728/next
2025-12-04T08:53:09.5516818Z  * [new branch]                gh/swolchok/819/base    -> origin/gh/swolchok/819/base
2025-12-04T08:53:09.5516889Z  * [new branch]                gh/swolchok/819/head    -> origin/gh/swolchok/819/head
2025-12-04T08:53:09.5516958Z  * [new branch]                gh/swolchok/819/orig    -> origin/gh/swolchok/819/orig
2025-12-04T08:53:09.5517030Z  * [new branch]                gh/swolchok/824/base    -> origin/gh/swolchok/824/base
2025-12-04T08:53:09.5517102Z  * [new branch]                gh/swolchok/824/head    -> origin/gh/swolchok/824/head
2025-12-04T08:53:09.5517175Z  * [new branch]                gh/swolchok/824/orig    -> origin/gh/swolchok/824/orig
2025-12-04T08:53:09.5517246Z  * [new branch]                gh/swolchok/829/base    -> origin/gh/swolchok/829/base
2025-12-04T08:53:09.5517318Z  * [new branch]                gh/swolchok/829/head    -> origin/gh/swolchok/829/head
2025-12-04T08:53:09.5517387Z  * [new branch]                gh/swolchok/829/orig    -> origin/gh/swolchok/829/orig
2025-12-04T08:53:09.5517459Z  * [new branch]                gh/swolchok/839/base    -> origin/gh/swolchok/839/base
2025-12-04T08:53:09.5517529Z  * [new branch]                gh/swolchok/839/head    -> origin/gh/swolchok/839/head
2025-12-04T08:53:09.5517600Z  * [new branch]                gh/swolchok/839/orig    -> origin/gh/swolchok/839/orig
2025-12-04T08:53:09.5517671Z  * [new branch]                gh/swolchok/841/base    -> origin/gh/swolchok/841/base
2025-12-04T08:53:09.5517742Z  * [new branch]                gh/swolchok/841/head    -> origin/gh/swolchok/841/head
2025-12-04T08:53:09.5517816Z  * [new branch]                gh/swolchok/841/orig    -> origin/gh/swolchok/841/orig
2025-12-04T08:53:09.5517886Z  * [new branch]                gh/swolchok/842/base    -> origin/gh/swolchok/842/base
2025-12-04T08:53:09.5517956Z  * [new branch]                gh/swolchok/842/head    -> origin/gh/swolchok/842/head
2025-12-04T08:53:09.5518026Z  * [new branch]                gh/swolchok/842/orig    -> origin/gh/swolchok/842/orig
2025-12-04T08:53:09.5518096Z  * [new branch]                gh/swolchok/845/base    -> origin/gh/swolchok/845/base
2025-12-04T08:53:09.5518167Z  * [new branch]                gh/swolchok/845/head    -> origin/gh/swolchok/845/head
2025-12-04T08:53:09.5518237Z  * [new branch]                gh/swolchok/845/orig    -> origin/gh/swolchok/845/orig
2025-12-04T08:53:09.5518307Z  * [new branch]                gh/swolchok/848/base    -> origin/gh/swolchok/848/base
2025-12-04T08:53:09.5518376Z  * [new branch]                gh/swolchok/848/head    -> origin/gh/swolchok/848/head
2025-12-04T08:53:09.5518448Z  * [new branch]                gh/swolchok/848/orig    -> origin/gh/swolchok/848/orig
2025-12-04T08:53:09.5518518Z  * [new branch]                gh/swolchok/856/base    -> origin/gh/swolchok/856/base
2025-12-04T08:53:09.5518589Z  * [new branch]                gh/swolchok/856/head    -> origin/gh/swolchok/856/head
2025-12-04T08:53:09.5518660Z  * [new branch]                gh/swolchok/856/orig    -> origin/gh/swolchok/856/orig
2025-12-04T08:53:09.5518729Z  * [new branch]                gh/swolchok/860/base    -> origin/gh/swolchok/860/base
2025-12-04T08:53:09.5518798Z  * [new branch]                gh/swolchok/860/head    -> origin/gh/swolchok/860/head
2025-12-04T08:53:09.5518868Z  * [new branch]                gh/swolchok/860/orig    -> origin/gh/swolchok/860/orig
2025-12-04T08:53:09.5518938Z  * [new branch]                gh/swolchok/861/base    -> origin/gh/swolchok/861/base
2025-12-04T08:53:09.5519008Z  * [new branch]                gh/swolchok/861/head    -> origin/gh/swolchok/861/head
2025-12-04T08:53:09.5519115Z  * [new branch]                gh/swolchok/861/orig    -> origin/gh/swolchok/861/orig
2025-12-04T08:53:09.5519216Z  * [new branch]                gh/swolchok/862/base    -> origin/gh/swolchok/862/base
2025-12-04T08:53:09.5519287Z  * [new branch]                gh/swolchok/862/head    -> origin/gh/swolchok/862/head
2025-12-04T08:53:09.5519358Z  * [new branch]                gh/swolchok/862/orig    -> origin/gh/swolchok/862/orig
2025-12-04T08:53:09.5519427Z  * [new branch]                gh/swolchok/863/base    -> origin/gh/swolchok/863/base
2025-12-04T08:53:09.5519498Z  * [new branch]                gh/swolchok/863/head    -> origin/gh/swolchok/863/head
2025-12-04T08:53:09.5519568Z  * [new branch]                gh/swolchok/863/orig    -> origin/gh/swolchok/863/orig
2025-12-04T08:53:09.5519638Z  * [new branch]                gh/swolchok/864/base    -> origin/gh/swolchok/864/base
2025-12-04T08:53:09.5519708Z  * [new branch]                gh/swolchok/864/head    -> origin/gh/swolchok/864/head
2025-12-04T08:53:09.5519785Z  * [new branch]                gh/swolchok/864/orig    -> origin/gh/swolchok/864/orig
2025-12-04T08:53:09.5519856Z  * [new branch]                gh/swolchok/865/base    -> origin/gh/swolchok/865/base
2025-12-04T08:53:09.5519927Z  * [new branch]                gh/swolchok/865/head    -> origin/gh/swolchok/865/head
2025-12-04T08:53:09.5519997Z  * [new branch]                gh/swolchok/865/orig    -> origin/gh/swolchok/865/orig
2025-12-04T08:53:09.5520066Z  * [new branch]                gh/swolchok/866/base    -> origin/gh/swolchok/866/base
2025-12-04T08:53:09.5520137Z  * [new branch]                gh/swolchok/866/head    -> origin/gh/swolchok/866/head
2025-12-04T08:53:09.5520206Z  * [new branch]                gh/swolchok/866/orig    -> origin/gh/swolchok/866/orig
2025-12-04T08:53:09.5520276Z  * [new branch]                gh/swolchok/867/base    -> origin/gh/swolchok/867/base
2025-12-04T08:53:09.5520346Z  * [new branch]                gh/swolchok/867/head    -> origin/gh/swolchok/867/head
2025-12-04T08:53:09.5520453Z  * [new branch]                gh/swolchok/867/orig    -> origin/gh/swolchok/867/orig
2025-12-04T08:53:09.5520524Z  * [new branch]                gh/swolchok/868/base    -> origin/gh/swolchok/868/base
2025-12-04T08:53:09.5520595Z  * [new branch]                gh/swolchok/868/head    -> origin/gh/swolchok/868/head
2025-12-04T08:53:09.5520665Z  * [new branch]                gh/swolchok/868/orig    -> origin/gh/swolchok/868/orig
2025-12-04T08:53:09.5520736Z  * [new branch]                gh/swolchok/869/base    -> origin/gh/swolchok/869/base
2025-12-04T08:53:09.5520805Z  * [new branch]                gh/swolchok/869/head    -> origin/gh/swolchok/869/head
2025-12-04T08:53:09.5520874Z  * [new branch]                gh/swolchok/869/orig    -> origin/gh/swolchok/869/orig
2025-12-04T08:53:09.5520945Z  * [new branch]                gh/swolchok/870/base    -> origin/gh/swolchok/870/base
2025-12-04T08:53:09.5521014Z  * [new branch]                gh/swolchok/870/head    -> origin/gh/swolchok/870/head
2025-12-04T08:53:09.5521086Z  * [new branch]                gh/swolchok/870/orig    -> origin/gh/swolchok/870/orig
2025-12-04T08:53:09.5521158Z  * [new branch]                gh/swolchok/871/base    -> origin/gh/swolchok/871/base
2025-12-04T08:53:09.5521229Z  * [new branch]                gh/swolchok/871/head    -> origin/gh/swolchok/871/head
2025-12-04T08:53:09.5521299Z  * [new branch]                gh/swolchok/871/orig    -> origin/gh/swolchok/871/orig
2025-12-04T08:53:09.5521372Z  * [new branch]                gh/teja-rao/4/base      -> origin/gh/teja-rao/4/base
2025-12-04T08:53:09.5521443Z  * [new branch]                gh/teja-rao/4/head      -> origin/gh/teja-rao/4/head
2025-12-04T08:53:09.5521511Z  * [new branch]                gh/teja-rao/4/orig      -> origin/gh/teja-rao/4/orig
2025-12-04T08:53:09.5521581Z  * [new branch]                gh/tianyu-l/2/base      -> origin/gh/tianyu-l/2/base
2025-12-04T08:53:09.5521648Z  * [new branch]                gh/tianyu-l/2/head      -> origin/gh/tianyu-l/2/head
2025-12-04T08:53:09.5521762Z  * [new branch]                gh/tianyu-l/2/orig      -> origin/gh/tianyu-l/2/orig
2025-12-04T08:53:09.5521874Z  * [new branch]                gh/tianyu-l/3/base      -> origin/gh/tianyu-l/3/base
2025-12-04T08:53:09.5521942Z  * [new branch]                gh/tianyu-l/3/orig      -> origin/gh/tianyu-l/3/orig
2025-12-04T08:53:09.5522008Z  * [new branch]                gh/tianyu-l/4/base      -> origin/gh/tianyu-l/4/base
2025-12-04T08:53:09.5522075Z  * [new branch]                gh/tianyu-l/4/head      -> origin/gh/tianyu-l/4/head
2025-12-04T08:53:09.5522141Z  * [new branch]                gh/tianyu-l/4/orig      -> origin/gh/tianyu-l/4/orig
2025-12-04T08:53:09.5522231Z  * [new branch]                gh/tugsbayasgalan/10/base -> origin/gh/tugsbayasgalan/10/base
2025-12-04T08:53:09.5522315Z  * [new branch]                gh/tugsbayasgalan/10/head -> origin/gh/tugsbayasgalan/10/head
2025-12-04T08:53:09.5522399Z  * [new branch]                gh/tugsbayasgalan/10/orig -> origin/gh/tugsbayasgalan/10/orig
2025-12-04T08:53:09.5522485Z  * [new branch]                gh/tugsbayasgalan/13/base -> origin/gh/tugsbayasgalan/13/base
2025-12-04T08:53:09.5522569Z  * [new branch]                gh/tugsbayasgalan/13/head -> origin/gh/tugsbayasgalan/13/head
2025-12-04T08:53:09.5522651Z  * [new branch]                gh/tugsbayasgalan/13/orig -> origin/gh/tugsbayasgalan/13/orig
2025-12-04T08:53:09.5522733Z  * [new branch]                gh/tugsbayasgalan/17/base -> origin/gh/tugsbayasgalan/17/base
2025-12-04T08:53:09.5522814Z  * [new branch]                gh/tugsbayasgalan/17/head -> origin/gh/tugsbayasgalan/17/head
2025-12-04T08:53:09.5522897Z  * [new branch]                gh/tugsbayasgalan/17/orig -> origin/gh/tugsbayasgalan/17/orig
2025-12-04T08:53:09.5522980Z  * [new branch]                gh/tugsbayasgalan/2/base -> origin/gh/tugsbayasgalan/2/base
2025-12-04T08:53:09.5523061Z  * [new branch]                gh/tugsbayasgalan/2/head -> origin/gh/tugsbayasgalan/2/head
2025-12-04T08:53:09.5523143Z  * [new branch]                gh/tugsbayasgalan/2/orig -> origin/gh/tugsbayasgalan/2/orig
2025-12-04T08:53:09.5523226Z  * [new branch]                gh/tugsbayasgalan/28/base -> origin/gh/tugsbayasgalan/28/base
2025-12-04T08:53:09.5523307Z  * [new branch]                gh/tugsbayasgalan/28/head -> origin/gh/tugsbayasgalan/28/head
2025-12-04T08:53:09.5523390Z  * [new branch]                gh/tugsbayasgalan/28/orig -> origin/gh/tugsbayasgalan/28/orig
2025-12-04T08:53:09.5523471Z  * [new branch]                gh/tugsbayasgalan/32/base -> origin/gh/tugsbayasgalan/32/base
2025-12-04T08:53:09.5523553Z  * [new branch]                gh/tugsbayasgalan/32/head -> origin/gh/tugsbayasgalan/32/head
2025-12-04T08:53:09.5523635Z  * [new branch]                gh/tugsbayasgalan/32/orig -> origin/gh/tugsbayasgalan/32/orig
2025-12-04T08:53:09.5523716Z  * [new branch]                gh/tugsbayasgalan/35/base -> origin/gh/tugsbayasgalan/35/base
2025-12-04T08:53:09.5523799Z  * [new branch]                gh/tugsbayasgalan/35/head -> origin/gh/tugsbayasgalan/35/head
2025-12-04T08:53:09.5523880Z  * [new branch]                gh/tugsbayasgalan/35/orig -> origin/gh/tugsbayasgalan/35/orig
2025-12-04T08:53:09.5523963Z  * [new branch]                gh/tugsbayasgalan/36/base -> origin/gh/tugsbayasgalan/36/base
2025-12-04T08:53:09.5524046Z  * [new branch]                gh/tugsbayasgalan/36/head -> origin/gh/tugsbayasgalan/36/head
2025-12-04T08:53:09.5524127Z  * [new branch]                gh/tugsbayasgalan/36/orig -> origin/gh/tugsbayasgalan/36/orig
2025-12-04T08:53:09.5524208Z  * [new branch]                gh/tugsbayasgalan/37/base -> origin/gh/tugsbayasgalan/37/base
2025-12-04T08:53:09.5524289Z  * [new branch]                gh/tugsbayasgalan/37/head -> origin/gh/tugsbayasgalan/37/head
2025-12-04T08:53:09.5524370Z  * [new branch]                gh/tugsbayasgalan/37/orig -> origin/gh/tugsbayasgalan/37/orig
2025-12-04T08:53:09.5524451Z  * [new branch]                gh/tugsbayasgalan/43/base -> origin/gh/tugsbayasgalan/43/base
2025-12-04T08:53:09.5524563Z  * [new branch]                gh/tugsbayasgalan/43/head -> origin/gh/tugsbayasgalan/43/head
2025-12-04T08:53:09.5524673Z  * [new branch]                gh/tugsbayasgalan/43/orig -> origin/gh/tugsbayasgalan/43/orig
2025-12-04T08:53:09.5524754Z  * [new branch]                gh/tugsbayasgalan/48/base -> origin/gh/tugsbayasgalan/48/base
2025-12-04T08:53:09.5524836Z  * [new branch]                gh/tugsbayasgalan/48/head -> origin/gh/tugsbayasgalan/48/head
2025-12-04T08:53:09.5524917Z  * [new branch]                gh/tugsbayasgalan/48/orig -> origin/gh/tugsbayasgalan/48/orig
2025-12-04T08:53:09.5525000Z  * [new branch]                gh/tugsbayasgalan/51/base -> origin/gh/tugsbayasgalan/51/base
2025-12-04T08:53:09.5525080Z  * [new branch]                gh/tugsbayasgalan/51/head -> origin/gh/tugsbayasgalan/51/head
2025-12-04T08:53:09.5525161Z  * [new branch]                gh/tugsbayasgalan/51/orig -> origin/gh/tugsbayasgalan/51/orig
2025-12-04T08:53:09.5525244Z  * [new branch]                gh/tugsbayasgalan/52/base -> origin/gh/tugsbayasgalan/52/base
2025-12-04T08:53:09.5525327Z  * [new branch]                gh/tugsbayasgalan/52/head -> origin/gh/tugsbayasgalan/52/head
2025-12-04T08:53:09.5525408Z  * [new branch]                gh/tugsbayasgalan/52/orig -> origin/gh/tugsbayasgalan/52/orig
2025-12-04T08:53:09.5525491Z  * [new branch]                gh/tugsbayasgalan/53/base -> origin/gh/tugsbayasgalan/53/base
2025-12-04T08:53:09.5525572Z  * [new branch]                gh/tugsbayasgalan/53/head -> origin/gh/tugsbayasgalan/53/head
2025-12-04T08:53:09.5525654Z  * [new branch]                gh/tugsbayasgalan/53/orig -> origin/gh/tugsbayasgalan/53/orig
2025-12-04T08:53:09.5525735Z  * [new branch]                gh/tugsbayasgalan/55/base -> origin/gh/tugsbayasgalan/55/base
2025-12-04T08:53:09.5525817Z  * [new branch]                gh/tugsbayasgalan/55/head -> origin/gh/tugsbayasgalan/55/head
2025-12-04T08:53:09.5525899Z  * [new branch]                gh/tugsbayasgalan/55/orig -> origin/gh/tugsbayasgalan/55/orig
2025-12-04T08:53:09.5525983Z  * [new branch]                gh/tugsbayasgalan/59/base -> origin/gh/tugsbayasgalan/59/base
2025-12-04T08:53:09.5526065Z  * [new branch]                gh/tugsbayasgalan/59/head -> origin/gh/tugsbayasgalan/59/head
2025-12-04T08:53:09.5526146Z  * [new branch]                gh/tugsbayasgalan/59/orig -> origin/gh/tugsbayasgalan/59/orig
2025-12-04T08:53:09.5526227Z  * [new branch]                gh/tugsbayasgalan/6/base -> origin/gh/tugsbayasgalan/6/base
2025-12-04T08:53:09.5526306Z  * [new branch]                gh/tugsbayasgalan/6/head -> origin/gh/tugsbayasgalan/6/head
2025-12-04T08:53:09.5526385Z  * [new branch]                gh/tugsbayasgalan/6/orig -> origin/gh/tugsbayasgalan/6/orig
2025-12-04T08:53:09.5526468Z  * [new branch]                gh/tugsbayasgalan/60/base -> origin/gh/tugsbayasgalan/60/base
2025-12-04T08:53:09.5526549Z  * [new branch]                gh/tugsbayasgalan/60/head -> origin/gh/tugsbayasgalan/60/head
2025-12-04T08:53:09.5526633Z  * [new branch]                gh/tugsbayasgalan/60/orig -> origin/gh/tugsbayasgalan/60/orig
2025-12-04T08:53:09.5526715Z  * [new branch]                gh/tugsbayasgalan/61/base -> origin/gh/tugsbayasgalan/61/base
2025-12-04T08:53:09.5526796Z  * [new branch]                gh/tugsbayasgalan/61/head -> origin/gh/tugsbayasgalan/61/head
2025-12-04T08:53:09.5526878Z  * [new branch]                gh/tugsbayasgalan/61/orig -> origin/gh/tugsbayasgalan/61/orig
2025-12-04T08:53:09.5526959Z  * [new branch]                gh/tugsbayasgalan/63/base -> origin/gh/tugsbayasgalan/63/base
2025-12-04T08:53:09.5527040Z  * [new branch]                gh/tugsbayasgalan/63/head -> origin/gh/tugsbayasgalan/63/head
2025-12-04T08:53:09.5527121Z  * [new branch]                gh/tugsbayasgalan/63/orig -> origin/gh/tugsbayasgalan/63/orig
2025-12-04T08:53:09.5527203Z  * [new branch]                gh/tugsbayasgalan/67/base -> origin/gh/tugsbayasgalan/67/base
2025-12-04T08:53:09.5527314Z  * [new branch]                gh/tugsbayasgalan/67/head -> origin/gh/tugsbayasgalan/67/head
2025-12-04T08:53:09.5527396Z  * [new branch]                gh/tugsbayasgalan/67/orig -> origin/gh/tugsbayasgalan/67/orig
2025-12-04T08:53:09.5527502Z  * [new branch]                gh/tugsbayasgalan/68/base -> origin/gh/tugsbayasgalan/68/base
2025-12-04T08:53:09.5527584Z  * [new branch]                gh/tugsbayasgalan/68/head -> origin/gh/tugsbayasgalan/68/head
2025-12-04T08:53:09.5527667Z  * [new branch]                gh/tugsbayasgalan/68/orig -> origin/gh/tugsbayasgalan/68/orig
2025-12-04T08:53:09.5527747Z  * [new branch]                gh/tugsbayasgalan/7/base -> origin/gh/tugsbayasgalan/7/base
2025-12-04T08:53:09.5527826Z  * [new branch]                gh/tugsbayasgalan/7/head -> origin/gh/tugsbayasgalan/7/head
2025-12-04T08:53:09.5527906Z  * [new branch]                gh/tugsbayasgalan/7/orig -> origin/gh/tugsbayasgalan/7/orig
2025-12-04T08:53:09.5527988Z  * [new branch]                gh/tugsbayasgalan/70/base -> origin/gh/tugsbayasgalan/70/base
2025-12-04T08:53:09.5528072Z  * [new branch]                gh/tugsbayasgalan/70/head -> origin/gh/tugsbayasgalan/70/head
2025-12-04T08:53:09.5528154Z  * [new branch]                gh/tugsbayasgalan/70/orig -> origin/gh/tugsbayasgalan/70/orig
2025-12-04T08:53:09.5528235Z  * [new branch]                gh/tugsbayasgalan/71/base -> origin/gh/tugsbayasgalan/71/base
2025-12-04T08:53:09.5528316Z  * [new branch]                gh/tugsbayasgalan/71/head -> origin/gh/tugsbayasgalan/71/head
2025-12-04T08:53:09.5528397Z  * [new branch]                gh/tugsbayasgalan/71/orig -> origin/gh/tugsbayasgalan/71/orig
2025-12-04T08:53:09.5528478Z  * [new branch]                gh/tugsbayasgalan/72/base -> origin/gh/tugsbayasgalan/72/base
2025-12-04T08:53:09.5528560Z  * [new branch]                gh/tugsbayasgalan/72/head -> origin/gh/tugsbayasgalan/72/head
2025-12-04T08:53:09.5528641Z  * [new branch]                gh/tugsbayasgalan/72/orig -> origin/gh/tugsbayasgalan/72/orig
2025-12-04T08:53:09.5528725Z  * [new branch]                gh/tugsbayasgalan/73/base -> origin/gh/tugsbayasgalan/73/base
2025-12-04T08:53:09.5528808Z  * [new branch]                gh/tugsbayasgalan/73/head -> origin/gh/tugsbayasgalan/73/head
2025-12-04T08:53:09.5528891Z  * [new branch]                gh/tugsbayasgalan/73/orig -> origin/gh/tugsbayasgalan/73/orig
2025-12-04T08:53:09.5528972Z  * [new branch]                gh/tugsbayasgalan/74/base -> origin/gh/tugsbayasgalan/74/base
2025-12-04T08:53:09.5529054Z  * [new branch]                gh/tugsbayasgalan/74/head -> origin/gh/tugsbayasgalan/74/head
2025-12-04T08:53:09.5529135Z  * [new branch]                gh/tugsbayasgalan/74/orig -> origin/gh/tugsbayasgalan/74/orig
2025-12-04T08:53:09.5529216Z  * [new branch]                gh/tugsbayasgalan/75/base -> origin/gh/tugsbayasgalan/75/base
2025-12-04T08:53:09.5529299Z  * [new branch]                gh/tugsbayasgalan/75/head -> origin/gh/tugsbayasgalan/75/head
2025-12-04T08:53:09.5529380Z  * [new branch]                gh/tugsbayasgalan/75/orig -> origin/gh/tugsbayasgalan/75/orig
2025-12-04T08:53:09.5529465Z  * [new branch]                gh/tugsbayasgalan/76/base -> origin/gh/tugsbayasgalan/76/base
2025-12-04T08:53:09.5529547Z  * [new branch]                gh/tugsbayasgalan/76/head -> origin/gh/tugsbayasgalan/76/head
2025-12-04T08:53:09.5529629Z  * [new branch]                gh/tugsbayasgalan/76/orig -> origin/gh/tugsbayasgalan/76/orig
2025-12-04T08:53:09.5529711Z  * [new branch]                gh/tugsbayasgalan/77/base -> origin/gh/tugsbayasgalan/77/base
2025-12-04T08:53:09.5529791Z  * [new branch]                gh/tugsbayasgalan/77/head -> origin/gh/tugsbayasgalan/77/head
2025-12-04T08:53:09.5529873Z  * [new branch]                gh/tugsbayasgalan/77/orig -> origin/gh/tugsbayasgalan/77/orig
2025-12-04T08:53:09.5529955Z  * [new branch]                gh/tugsbayasgalan/78/base -> origin/gh/tugsbayasgalan/78/base
2025-12-04T08:53:09.5530035Z  * [new branch]                gh/tugsbayasgalan/78/head -> origin/gh/tugsbayasgalan/78/head
2025-12-04T08:53:09.5530148Z  * [new branch]                gh/tugsbayasgalan/78/orig -> origin/gh/tugsbayasgalan/78/orig
2025-12-04T08:53:09.5530252Z  * [new branch]                gh/tugsbayasgalan/79/base -> origin/gh/tugsbayasgalan/79/base
2025-12-04T08:53:09.5530336Z  * [new branch]                gh/tugsbayasgalan/79/head -> origin/gh/tugsbayasgalan/79/head
2025-12-04T08:53:09.5530438Z  * [new branch]                gh/tugsbayasgalan/79/orig -> origin/gh/tugsbayasgalan/79/orig
2025-12-04T08:53:09.5530522Z  * [new branch]                gh/tugsbayasgalan/8/base -> origin/gh/tugsbayasgalan/8/base
2025-12-04T08:53:09.5530601Z  * [new branch]                gh/tugsbayasgalan/8/head -> origin/gh/tugsbayasgalan/8/head
2025-12-04T08:53:09.5530680Z  * [new branch]                gh/tugsbayasgalan/8/orig -> origin/gh/tugsbayasgalan/8/orig
2025-12-04T08:53:09.5530767Z  * [new branch]                gh/tugsbayasgalan/80/base -> origin/gh/tugsbayasgalan/80/base
2025-12-04T08:53:09.5530848Z  * [new branch]                gh/tugsbayasgalan/80/head -> origin/gh/tugsbayasgalan/80/head
2025-12-04T08:53:09.5530931Z  * [new branch]                gh/tugsbayasgalan/80/orig -> origin/gh/tugsbayasgalan/80/orig
2025-12-04T08:53:09.5531014Z  * [new branch]                gh/tugsbayasgalan/81/base -> origin/gh/tugsbayasgalan/81/base
2025-12-04T08:53:09.5531095Z  * [new branch]                gh/tugsbayasgalan/81/head -> origin/gh/tugsbayasgalan/81/head
2025-12-04T08:53:09.5531177Z  * [new branch]                gh/tugsbayasgalan/81/orig -> origin/gh/tugsbayasgalan/81/orig
2025-12-04T08:53:09.5531258Z  * [new branch]                gh/tugsbayasgalan/82/base -> origin/gh/tugsbayasgalan/82/base
2025-12-04T08:53:09.5531339Z  * [new branch]                gh/tugsbayasgalan/82/head -> origin/gh/tugsbayasgalan/82/head
2025-12-04T08:53:09.5531421Z  * [new branch]                gh/tugsbayasgalan/82/orig -> origin/gh/tugsbayasgalan/82/orig
2025-12-04T08:53:09.5531502Z  * [new branch]                gh/tugsbayasgalan/83/base -> origin/gh/tugsbayasgalan/83/base
2025-12-04T08:53:09.5531584Z  * [new branch]                gh/tugsbayasgalan/83/head -> origin/gh/tugsbayasgalan/83/head
2025-12-04T08:53:09.5531669Z  * [new branch]                gh/tugsbayasgalan/83/orig -> origin/gh/tugsbayasgalan/83/orig
2025-12-04T08:53:09.5531751Z  * [new branch]                gh/tugsbayasgalan/84/base -> origin/gh/tugsbayasgalan/84/base
2025-12-04T08:53:09.5531833Z  * [new branch]                gh/tugsbayasgalan/84/head -> origin/gh/tugsbayasgalan/84/head
2025-12-04T08:53:09.5531915Z  * [new branch]                gh/tugsbayasgalan/84/orig -> origin/gh/tugsbayasgalan/84/orig
2025-12-04T08:53:09.5531995Z  * [new branch]                gh/tugsbayasgalan/85/base -> origin/gh/tugsbayasgalan/85/base
2025-12-04T08:53:09.5532078Z  * [new branch]                gh/tugsbayasgalan/85/head -> origin/gh/tugsbayasgalan/85/head
2025-12-04T08:53:09.5532160Z  * [new branch]                gh/tugsbayasgalan/85/orig -> origin/gh/tugsbayasgalan/85/orig
2025-12-04T08:53:09.5532244Z  * [new branch]                gh/tugsbayasgalan/86/base -> origin/gh/tugsbayasgalan/86/base
2025-12-04T08:53:09.5532325Z  * [new branch]                gh/tugsbayasgalan/86/head -> origin/gh/tugsbayasgalan/86/head
2025-12-04T08:53:09.5532408Z  * [new branch]                gh/tugsbayasgalan/86/orig -> origin/gh/tugsbayasgalan/86/orig
2025-12-04T08:53:09.5532489Z  * [new branch]                gh/tugsbayasgalan/87/base -> origin/gh/tugsbayasgalan/87/base
2025-12-04T08:53:09.5532570Z  * [new branch]                gh/tugsbayasgalan/87/head -> origin/gh/tugsbayasgalan/87/head
2025-12-04T08:53:09.5532652Z  * [new branch]                gh/tugsbayasgalan/87/orig -> origin/gh/tugsbayasgalan/87/orig
2025-12-04T08:53:09.5532733Z  * [new branch]                gh/tugsbayasgalan/88/base -> origin/gh/tugsbayasgalan/88/base
2025-12-04T08:53:09.5532815Z  * [new branch]                gh/tugsbayasgalan/88/head -> origin/gh/tugsbayasgalan/88/head
2025-12-04T08:53:09.5532895Z  * [new branch]                gh/tugsbayasgalan/88/orig -> origin/gh/tugsbayasgalan/88/orig
2025-12-04T08:53:09.5533019Z  * [new branch]                gh/tugsbayasgalan/89/base -> origin/gh/tugsbayasgalan/89/base
2025-12-04T08:53:09.5533144Z  * [new branch]                gh/tugsbayasgalan/89/head -> origin/gh/tugsbayasgalan/89/head
2025-12-04T08:53:09.5533225Z  * [new branch]                gh/tugsbayasgalan/89/orig -> origin/gh/tugsbayasgalan/89/orig
2025-12-04T08:53:09.5533304Z  * [new branch]                gh/tugsbayasgalan/9/base -> origin/gh/tugsbayasgalan/9/base
2025-12-04T08:53:09.5533385Z  * [new branch]                gh/tugsbayasgalan/9/head -> origin/gh/tugsbayasgalan/9/head
2025-12-04T08:53:09.5533464Z  * [new branch]                gh/tugsbayasgalan/9/orig -> origin/gh/tugsbayasgalan/9/orig
2025-12-04T08:53:09.5533546Z  * [new branch]                gh/tugsbayasgalan/90/base -> origin/gh/tugsbayasgalan/90/base
2025-12-04T08:53:09.5533630Z  * [new branch]                gh/tugsbayasgalan/90/head -> origin/gh/tugsbayasgalan/90/head
2025-12-04T08:53:09.5533713Z  * [new branch]                gh/tugsbayasgalan/90/orig -> origin/gh/tugsbayasgalan/90/orig
2025-12-04T08:53:09.5533796Z  * [new branch]                gh/tugsbayasgalan/91/base -> origin/gh/tugsbayasgalan/91/base
2025-12-04T08:53:09.5533879Z  * [new branch]                gh/tugsbayasgalan/91/head -> origin/gh/tugsbayasgalan/91/head
2025-12-04T08:53:09.5533959Z  * [new branch]                gh/tugsbayasgalan/91/orig -> origin/gh/tugsbayasgalan/91/orig
2025-12-04T08:53:09.5534040Z  * [new branch]                gh/tugsbayasgalan/92/base -> origin/gh/tugsbayasgalan/92/base
2025-12-04T08:53:09.5534123Z  * [new branch]                gh/tugsbayasgalan/92/head -> origin/gh/tugsbayasgalan/92/head
2025-12-04T08:53:09.5534204Z  * [new branch]                gh/tugsbayasgalan/92/orig -> origin/gh/tugsbayasgalan/92/orig
2025-12-04T08:53:09.5534285Z  * [new branch]                gh/tugsbayasgalan/93/base -> origin/gh/tugsbayasgalan/93/base
2025-12-04T08:53:09.5534368Z  * [new branch]                gh/tugsbayasgalan/93/head -> origin/gh/tugsbayasgalan/93/head
2025-12-04T08:53:09.5534451Z  * [new branch]                gh/tugsbayasgalan/93/orig -> origin/gh/tugsbayasgalan/93/orig
2025-12-04T08:53:09.5534520Z  * [new branch]                gh/v0i0/14/base         -> origin/gh/v0i0/14/base
2025-12-04T08:53:09.5534585Z  * [new branch]                gh/v0i0/14/head         -> origin/gh/v0i0/14/head
2025-12-04T08:53:09.5534649Z  * [new branch]                gh/v0i0/14/orig         -> origin/gh/v0i0/14/orig
2025-12-04T08:53:09.5534713Z  * [new branch]                gh/v0i0/15/base         -> origin/gh/v0i0/15/base
2025-12-04T08:53:09.5534775Z  * [new branch]                gh/v0i0/15/head         -> origin/gh/v0i0/15/head
2025-12-04T08:53:09.5534836Z  * [new branch]                gh/v0i0/15/orig         -> origin/gh/v0i0/15/orig
2025-12-04T08:53:09.5534898Z  * [new branch]                gh/v0i0/16/base         -> origin/gh/v0i0/16/base
2025-12-04T08:53:09.5534960Z  * [new branch]                gh/v0i0/16/head         -> origin/gh/v0i0/16/head
2025-12-04T08:53:09.5535024Z  * [new branch]                gh/v0i0/16/orig         -> origin/gh/v0i0/16/orig
2025-12-04T08:53:09.5535088Z  * [new branch]                gh/v0i0/17/base         -> origin/gh/v0i0/17/base
2025-12-04T08:53:09.5535150Z  * [new branch]                gh/v0i0/17/head         -> origin/gh/v0i0/17/head
2025-12-04T08:53:09.5535211Z  * [new branch]                gh/v0i0/17/orig         -> origin/gh/v0i0/17/orig
2025-12-04T08:53:09.5535273Z  * [new branch]                gh/v0i0/18/base         -> origin/gh/v0i0/18/base
2025-12-04T08:53:09.5535334Z  * [new branch]                gh/v0i0/18/head         -> origin/gh/v0i0/18/head
2025-12-04T08:53:09.5535396Z  * [new branch]                gh/v0i0/18/orig         -> origin/gh/v0i0/18/orig
2025-12-04T08:53:09.5535458Z  * [new branch]                gh/v0i0/19/base         -> origin/gh/v0i0/19/base
2025-12-04T08:53:09.5535519Z  * [new branch]                gh/v0i0/19/head         -> origin/gh/v0i0/19/head
2025-12-04T08:53:09.5535606Z  * [new branch]                gh/v0i0/19/orig         -> origin/gh/v0i0/19/orig
2025-12-04T08:53:09.5535707Z  * [new branch]                gh/vishal9-team/1/base  -> origin/gh/vishal9-team/1/base
2025-12-04T08:53:09.5535783Z  * [new branch]                gh/vishal9-team/1/head  -> origin/gh/vishal9-team/1/head
2025-12-04T08:53:09.5535857Z  * [new branch]                gh/vishal9-team/2/base  -> origin/gh/vishal9-team/2/base
2025-12-04T08:53:09.5535932Z  * [new branch]                gh/vishal9-team/2/head  -> origin/gh/vishal9-team/2/head
2025-12-04T08:53:09.5536006Z  * [new branch]                gh/vishal9-team/2/orig  -> origin/gh/vishal9-team/2/orig
2025-12-04T08:53:09.5536080Z  * [new branch]                gh/vishal9-team/3/base  -> origin/gh/vishal9-team/3/base
2025-12-04T08:53:09.5536153Z  * [new branch]                gh/vishal9-team/3/head  -> origin/gh/vishal9-team/3/head
2025-12-04T08:53:09.5536226Z  * [new branch]                gh/vishal9-team/3/orig  -> origin/gh/vishal9-team/3/orig
2025-12-04T08:53:09.5536301Z  * [new branch]                gh/vishal9-team/4/base  -> origin/gh/vishal9-team/4/base
2025-12-04T08:53:09.5536376Z  * [new branch]                gh/vishal9-team/4/head  -> origin/gh/vishal9-team/4/head
2025-12-04T08:53:09.5536449Z  * [new branch]                gh/vishal9-team/4/orig  -> origin/gh/vishal9-team/4/orig
2025-12-04T08:53:09.5536515Z  * [new branch]                gh/vkuzo/1/next         -> origin/gh/vkuzo/1/next
2025-12-04T08:53:09.5536579Z  * [new branch]                gh/vkuzo/2/next         -> origin/gh/vkuzo/2/next
2025-12-04T08:53:09.5536644Z  * [new branch]                gh/vkuzo/3/next         -> origin/gh/vkuzo/3/next
2025-12-04T08:53:09.5536718Z  * [new branch]                gh/wconstab/424/base    -> origin/gh/wconstab/424/base
2025-12-04T08:53:09.5536791Z  * [new branch]                gh/wconstab/424/head    -> origin/gh/wconstab/424/head
2025-12-04T08:53:09.5536862Z  * [new branch]                gh/wconstab/424/orig    -> origin/gh/wconstab/424/orig
2025-12-04T08:53:09.5536935Z  * [new branch]                gh/wconstab/435/base    -> origin/gh/wconstab/435/base
2025-12-04T08:53:09.5537007Z  * [new branch]                gh/wconstab/435/head    -> origin/gh/wconstab/435/head
2025-12-04T08:53:09.5537077Z  * [new branch]                gh/wconstab/435/orig    -> origin/gh/wconstab/435/orig
2025-12-04T08:53:09.5537148Z  * [new branch]                gh/wconstab/444/base    -> origin/gh/wconstab/444/base
2025-12-04T08:53:09.5537218Z  * [new branch]                gh/wconstab/444/head    -> origin/gh/wconstab/444/head
2025-12-04T08:53:09.5537288Z  * [new branch]                gh/wconstab/444/orig    -> origin/gh/wconstab/444/orig
2025-12-04T08:53:09.5537359Z  * [new branch]                gh/wconstab/447/base    -> origin/gh/wconstab/447/base
2025-12-04T08:53:09.5537428Z  * [new branch]                gh/wconstab/447/head    -> origin/gh/wconstab/447/head
2025-12-04T08:53:09.5537498Z  * [new branch]                gh/wconstab/447/orig    -> origin/gh/wconstab/447/orig
2025-12-04T08:53:09.5537571Z  * [new branch]                gh/wconstab/448/base    -> origin/gh/wconstab/448/base
2025-12-04T08:53:09.5537642Z  * [new branch]                gh/wconstab/448/head    -> origin/gh/wconstab/448/head
2025-12-04T08:53:09.5537715Z  * [new branch]                gh/wconstab/448/orig    -> origin/gh/wconstab/448/orig
2025-12-04T08:53:09.5537785Z  * [new branch]                gh/wconstab/449/base    -> origin/gh/wconstab/449/base
2025-12-04T08:53:09.5537855Z  * [new branch]                gh/wconstab/449/head    -> origin/gh/wconstab/449/head
2025-12-04T08:53:09.5537925Z  * [new branch]                gh/wconstab/449/orig    -> origin/gh/wconstab/449/orig
2025-12-04T08:53:09.5537995Z  * [new branch]                gh/wconstab/450/base    -> origin/gh/wconstab/450/base
2025-12-04T08:53:09.5538064Z  * [new branch]                gh/wconstab/450/head    -> origin/gh/wconstab/450/head
2025-12-04T08:53:09.5538136Z  * [new branch]                gh/wconstab/450/orig    -> origin/gh/wconstab/450/orig
2025-12-04T08:53:09.5538230Z  * [new branch]                gh/wconstab/451/base    -> origin/gh/wconstab/451/base
2025-12-04T08:53:09.5538335Z  * [new branch]                gh/wconstab/451/head    -> origin/gh/wconstab/451/head
2025-12-04T08:53:09.5538406Z  * [new branch]                gh/wconstab/451/orig    -> origin/gh/wconstab/451/orig
2025-12-04T08:53:09.5538476Z  * [new branch]                gh/wconstab/452/base    -> origin/gh/wconstab/452/base
2025-12-04T08:53:09.5538546Z  * [new branch]                gh/wconstab/452/head    -> origin/gh/wconstab/452/head
2025-12-04T08:53:09.5538618Z  * [new branch]                gh/wconstab/452/orig    -> origin/gh/wconstab/452/orig
2025-12-04T08:53:09.5538687Z  * [new branch]                gh/wconstab/453/base    -> origin/gh/wconstab/453/base
2025-12-04T08:53:09.5538756Z  * [new branch]                gh/wconstab/453/head    -> origin/gh/wconstab/453/head
2025-12-04T08:53:09.5538830Z  * [new branch]                gh/wconstab/453/orig    -> origin/gh/wconstab/453/orig
2025-12-04T08:53:09.5538900Z  * [new branch]                gh/wconstab/454/base    -> origin/gh/wconstab/454/base
2025-12-04T08:53:09.5538970Z  * [new branch]                gh/wconstab/454/head    -> origin/gh/wconstab/454/head
2025-12-04T08:53:09.5539041Z  * [new branch]                gh/wconstab/454/orig    -> origin/gh/wconstab/454/orig
2025-12-04T08:53:09.5539110Z  * [new branch]                gh/wconstab/455/base    -> origin/gh/wconstab/455/base
2025-12-04T08:53:09.5539184Z  * [new branch]                gh/wconstab/455/head    -> origin/gh/wconstab/455/head
2025-12-04T08:53:09.5539254Z  * [new branch]                gh/wconstab/455/orig    -> origin/gh/wconstab/455/orig
2025-12-04T08:53:09.5539323Z  * [new branch]                gh/wconstab/456/base    -> origin/gh/wconstab/456/base
2025-12-04T08:53:09.5539396Z  * [new branch]                gh/wconstab/456/head    -> origin/gh/wconstab/456/head
2025-12-04T08:53:09.5539467Z  * [new branch]                gh/wconstab/456/orig    -> origin/gh/wconstab/456/orig
2025-12-04T08:53:09.5539538Z  * [new branch]                gh/wconstab/457/base    -> origin/gh/wconstab/457/base
2025-12-04T08:53:09.5539611Z  * [new branch]                gh/wconstab/457/head    -> origin/gh/wconstab/457/head
2025-12-04T08:53:09.5539681Z  * [new branch]                gh/wconstab/457/orig    -> origin/gh/wconstab/457/orig
2025-12-04T08:53:09.5539750Z  * [new branch]                gh/wconstab/458/base    -> origin/gh/wconstab/458/base
2025-12-04T08:53:09.5539820Z  * [new branch]                gh/wconstab/458/head    -> origin/gh/wconstab/458/head
2025-12-04T08:53:09.5539889Z  * [new branch]                gh/wconstab/458/orig    -> origin/gh/wconstab/458/orig
2025-12-04T08:53:09.5539959Z  * [new branch]                gh/wconstab/459/base    -> origin/gh/wconstab/459/base
2025-12-04T08:53:09.5540031Z  * [new branch]                gh/wconstab/459/head    -> origin/gh/wconstab/459/head
2025-12-04T08:53:09.5540102Z  * [new branch]                gh/wconstab/459/orig    -> origin/gh/wconstab/459/orig
2025-12-04T08:53:09.5540171Z  * [new branch]                gh/wconstab/460/base    -> origin/gh/wconstab/460/base
2025-12-04T08:53:09.5540244Z  * [new branch]                gh/wconstab/460/head    -> origin/gh/wconstab/460/head
2025-12-04T08:53:09.5540315Z  * [new branch]                gh/wconstab/460/orig    -> origin/gh/wconstab/460/orig
2025-12-04T08:53:09.5540384Z  * [new branch]                gh/wconstab/461/base    -> origin/gh/wconstab/461/base
2025-12-04T08:53:09.5540489Z  * [new branch]                gh/wconstab/461/head    -> origin/gh/wconstab/461/head
2025-12-04T08:53:09.5540559Z  * [new branch]                gh/wconstab/461/orig    -> origin/gh/wconstab/461/orig
2025-12-04T08:53:09.5540636Z  * [new branch]                gh/wconstab/462/base    -> origin/gh/wconstab/462/base
2025-12-04T08:53:09.5540705Z  * [new branch]                gh/wconstab/462/head    -> origin/gh/wconstab/462/head
2025-12-04T08:53:09.5540817Z  * [new branch]                gh/wconstab/462/orig    -> origin/gh/wconstab/462/orig
2025-12-04T08:53:09.5540889Z  * [new branch]                gh/wconstab/463/base    -> origin/gh/wconstab/463/base
2025-12-04T08:53:09.5541004Z  * [new branch]                gh/wconstab/463/head    -> origin/gh/wconstab/463/head
2025-12-04T08:53:09.5541075Z  * [new branch]                gh/wconstab/463/orig    -> origin/gh/wconstab/463/orig
2025-12-04T08:53:09.5541145Z  * [new branch]                gh/wconstab/464/base    -> origin/gh/wconstab/464/base
2025-12-04T08:53:09.5541215Z  * [new branch]                gh/wconstab/464/head    -> origin/gh/wconstab/464/head
2025-12-04T08:53:09.5541286Z  * [new branch]                gh/wconstab/464/orig    -> origin/gh/wconstab/464/orig
2025-12-04T08:53:09.5541366Z  * [new branch]                gh/wconstab/465/base    -> origin/gh/wconstab/465/base
2025-12-04T08:53:09.5541567Z  * [new branch]                gh/wconstab/465/head    -> origin/gh/wconstab/465/head
2025-12-04T08:53:09.5541770Z  * [new branch]                gh/wconstab/465/orig    -> origin/gh/wconstab/465/orig
2025-12-04T08:53:09.5541970Z  * [new branch]                gh/wconstab/466/base    -> origin/gh/wconstab/466/base
2025-12-04T08:53:09.5542153Z  * [new branch]                gh/wconstab/466/head    -> origin/gh/wconstab/466/head
2025-12-04T08:53:09.5542336Z  * [new branch]                gh/wconstab/466/orig    -> origin/gh/wconstab/466/orig
2025-12-04T08:53:09.5542519Z  * [new branch]                gh/wconstab/467/base    -> origin/gh/wconstab/467/base
2025-12-04T08:53:09.5542700Z  * [new branch]                gh/wconstab/467/head    -> origin/gh/wconstab/467/head
2025-12-04T08:53:09.5542882Z  * [new branch]                gh/wconstab/467/orig    -> origin/gh/wconstab/467/orig
2025-12-04T08:53:09.5543060Z  * [new branch]                gh/wconstab/468/base    -> origin/gh/wconstab/468/base
2025-12-04T08:53:09.5543248Z  * [new branch]                gh/wconstab/468/head    -> origin/gh/wconstab/468/head
2025-12-04T08:53:09.5543428Z  * [new branch]                gh/wconstab/468/orig    -> origin/gh/wconstab/468/orig
2025-12-04T08:53:09.5543610Z  * [new branch]                gh/weifengpy/39/base    -> origin/gh/weifengpy/39/base
2025-12-04T08:53:09.5543808Z  * [new branch]                gh/weifengpy/39/head    -> origin/gh/weifengpy/39/head
2025-12-04T08:53:09.5543991Z  * [new branch]                gh/weifengpy/39/orig    -> origin/gh/weifengpy/39/orig
2025-12-04T08:53:09.5544171Z  * [new branch]                gh/weifengpy/40/base    -> origin/gh/weifengpy/40/base
2025-12-04T08:53:09.5544351Z  * [new branch]                gh/weifengpy/40/head    -> origin/gh/weifengpy/40/head
2025-12-04T08:53:09.5544533Z  * [new branch]                gh/weifengpy/40/orig    -> origin/gh/weifengpy/40/orig
2025-12-04T08:53:09.5544716Z  * [new branch]                gh/weifengpy/41/base    -> origin/gh/weifengpy/41/base
2025-12-04T08:53:09.5544900Z  * [new branch]                gh/weifengpy/41/head    -> origin/gh/weifengpy/41/head
2025-12-04T08:53:09.5545081Z  * [new branch]                gh/weifengpy/41/orig    -> origin/gh/weifengpy/41/orig
2025-12-04T08:53:09.5545273Z  * [new branch]                gh/williamwen42/250/base -> origin/gh/williamwen42/250/base
2025-12-04T08:53:09.5545472Z  * [new branch]                gh/williamwen42/250/head -> origin/gh/williamwen42/250/head
2025-12-04T08:53:09.5545673Z  * [new branch]                gh/williamwen42/250/orig -> origin/gh/williamwen42/250/orig
2025-12-04T08:53:09.5545869Z  * [new branch]                gh/williamwen42/279/base -> origin/gh/williamwen42/279/base
2025-12-04T08:53:09.5546060Z  * [new branch]                gh/williamwen42/279/head -> origin/gh/williamwen42/279/head
2025-12-04T08:53:09.5546253Z  * [new branch]                gh/williamwen42/279/orig -> origin/gh/williamwen42/279/orig
2025-12-04T08:53:09.5546445Z  * [new branch]                gh/williamwen42/282/base -> origin/gh/williamwen42/282/base
2025-12-04T08:53:09.5546666Z  * [new branch]                gh/williamwen42/282/head -> origin/gh/williamwen42/282/head
2025-12-04T08:53:09.5546866Z  * [new branch]                gh/williamwen42/282/orig -> origin/gh/williamwen42/282/orig
2025-12-04T08:53:09.5547102Z  * [new branch]                gh/williamwen42/287/base -> origin/gh/williamwen42/287/base
2025-12-04T08:53:09.5547311Z  * [new branch]                gh/williamwen42/287/head -> origin/gh/williamwen42/287/head
2025-12-04T08:53:09.5547506Z  * [new branch]                gh/williamwen42/287/orig -> origin/gh/williamwen42/287/orig
2025-12-04T08:53:09.5547704Z  * [new branch]                gh/williamwen42/288/base -> origin/gh/williamwen42/288/base
2025-12-04T08:53:09.5547898Z  * [new branch]                gh/williamwen42/288/head -> origin/gh/williamwen42/288/head
2025-12-04T08:53:09.5548098Z  * [new branch]                gh/williamwen42/288/orig -> origin/gh/williamwen42/288/orig
2025-12-04T08:53:09.5548301Z  * [new branch]                gh/williamwen42/296/base -> origin/gh/williamwen42/296/base
2025-12-04T08:53:09.5548494Z  * [new branch]                gh/williamwen42/296/head -> origin/gh/williamwen42/296/head
2025-12-04T08:53:09.5548692Z  * [new branch]                gh/williamwen42/296/orig -> origin/gh/williamwen42/296/orig
2025-12-04T08:53:09.5548890Z  * [new branch]                gh/williamwen42/297/base -> origin/gh/williamwen42/297/base
2025-12-04T08:53:09.5549083Z  * [new branch]                gh/williamwen42/297/head -> origin/gh/williamwen42/297/head
2025-12-04T08:53:09.5549276Z  * [new branch]                gh/williamwen42/297/orig -> origin/gh/williamwen42/297/orig
2025-12-04T08:53:09.5549474Z  * [new branch]                gh/williamwen42/306/base -> origin/gh/williamwen42/306/base
2025-12-04T08:53:09.5549672Z  * [new branch]                gh/williamwen42/306/head -> origin/gh/williamwen42/306/head
2025-12-04T08:53:09.5549863Z  * [new branch]                gh/williamwen42/306/orig -> origin/gh/williamwen42/306/orig
2025-12-04T08:53:09.5550054Z  * [new branch]                gh/williamwen42/309/base -> origin/gh/williamwen42/309/base
2025-12-04T08:53:09.5550252Z  * [new branch]                gh/williamwen42/309/head -> origin/gh/williamwen42/309/head
2025-12-04T08:53:09.5550582Z  * [new branch]                gh/williamwen42/309/orig -> origin/gh/williamwen42/309/orig
2025-12-04T08:53:09.5550846Z  * [new branch]                gh/williamwen42/310/base -> origin/gh/williamwen42/310/base
2025-12-04T08:53:09.5551089Z  * [new branch]                gh/williamwen42/310/head -> origin/gh/williamwen42/310/head
2025-12-04T08:53:09.5551283Z  * [new branch]                gh/williamwen42/310/orig -> origin/gh/williamwen42/310/orig
2025-12-04T08:53:09.5551475Z  * [new branch]                gh/williamwen42/311/base -> origin/gh/williamwen42/311/base
2025-12-04T08:53:09.5551672Z  * [new branch]                gh/williamwen42/311/head -> origin/gh/williamwen42/311/head
2025-12-04T08:53:09.5551865Z  * [new branch]                gh/williamwen42/311/orig -> origin/gh/williamwen42/311/orig
2025-12-04T08:53:09.5552082Z  * [new branch]                gh/williamwen42/319/base -> origin/gh/williamwen42/319/base
2025-12-04T08:53:09.5552288Z  * [new branch]                gh/williamwen42/319/head -> origin/gh/williamwen42/319/head
2025-12-04T08:53:09.5552481Z  * [new branch]                gh/williamwen42/319/orig -> origin/gh/williamwen42/319/orig
2025-12-04T08:53:09.5552672Z  * [new branch]                gh/williamwen42/325/base -> origin/gh/williamwen42/325/base
2025-12-04T08:53:09.5552863Z  * [new branch]                gh/williamwen42/325/head -> origin/gh/williamwen42/325/head
2025-12-04T08:53:09.5553054Z  * [new branch]                gh/williamwen42/325/orig -> origin/gh/williamwen42/325/orig
2025-12-04T08:53:09.5553247Z  * [new branch]                gh/williamwen42/326/base -> origin/gh/williamwen42/326/base
2025-12-04T08:53:09.5553437Z  * [new branch]                gh/williamwen42/326/head -> origin/gh/williamwen42/326/head
2025-12-04T08:53:09.5553628Z  * [new branch]                gh/williamwen42/326/orig -> origin/gh/williamwen42/326/orig
2025-12-04T08:53:09.5553878Z  * [new branch]                gh/williamwen42/327/base -> origin/gh/williamwen42/327/base
2025-12-04T08:53:09.5554113Z  * [new branch]                gh/williamwen42/327/head -> origin/gh/williamwen42/327/head
2025-12-04T08:53:09.5554304Z  * [new branch]                gh/williamwen42/327/orig -> origin/gh/williamwen42/327/orig
2025-12-04T08:53:09.5554496Z  * [new branch]                gh/williamwen42/328/base -> origin/gh/williamwen42/328/base
2025-12-04T08:53:09.5554690Z  * [new branch]                gh/williamwen42/328/head -> origin/gh/williamwen42/328/head
2025-12-04T08:53:09.5554884Z  * [new branch]                gh/williamwen42/328/orig -> origin/gh/williamwen42/328/orig
2025-12-04T08:53:09.5555077Z  * [new branch]                gh/williamwen42/329/base -> origin/gh/williamwen42/329/base
2025-12-04T08:53:09.5555269Z  * [new branch]                gh/williamwen42/329/head -> origin/gh/williamwen42/329/head
2025-12-04T08:53:09.5555467Z  * [new branch]                gh/williamwen42/329/orig -> origin/gh/williamwen42/329/orig
2025-12-04T08:53:09.5555663Z  * [new branch]                gh/williamwen42/330/base -> origin/gh/williamwen42/330/base
2025-12-04T08:53:09.5555854Z  * [new branch]                gh/williamwen42/330/head -> origin/gh/williamwen42/330/head
2025-12-04T08:53:09.5556049Z  * [new branch]                gh/williamwen42/330/orig -> origin/gh/williamwen42/330/orig
2025-12-04T08:53:09.5556241Z  * [new branch]                gh/williamwen42/331/base -> origin/gh/williamwen42/331/base
2025-12-04T08:53:09.5556433Z  * [new branch]                gh/williamwen42/331/head -> origin/gh/williamwen42/331/head
2025-12-04T08:53:09.5556625Z  * [new branch]                gh/williamwen42/331/orig -> origin/gh/williamwen42/331/orig
2025-12-04T08:53:09.5556817Z  * [new branch]                gh/williamwen42/332/base -> origin/gh/williamwen42/332/base
2025-12-04T08:53:09.5557010Z  * [new branch]                gh/williamwen42/332/head -> origin/gh/williamwen42/332/head
2025-12-04T08:53:09.5557204Z  * [new branch]                gh/williamwen42/332/orig -> origin/gh/williamwen42/332/orig
2025-12-04T08:53:09.5557403Z  * [new branch]                gh/williamwen42/333/base -> origin/gh/williamwen42/333/base
2025-12-04T08:53:09.5557596Z  * [new branch]                gh/williamwen42/333/head -> origin/gh/williamwen42/333/head
2025-12-04T08:53:09.5557787Z  * [new branch]                gh/williamwen42/333/orig -> origin/gh/williamwen42/333/orig
2025-12-04T08:53:09.5557979Z  * [new branch]                gh/williamwen42/334/base -> origin/gh/williamwen42/334/base
2025-12-04T08:53:09.5558172Z  * [new branch]                gh/williamwen42/334/head -> origin/gh/williamwen42/334/head
2025-12-04T08:53:09.5558363Z  * [new branch]                gh/williamwen42/334/orig -> origin/gh/williamwen42/334/orig
2025-12-04T08:53:09.5558557Z  * [new branch]                gh/williamwen42/335/base -> origin/gh/williamwen42/335/base
2025-12-04T08:53:09.5558762Z  * [new branch]                gh/williamwen42/335/head -> origin/gh/williamwen42/335/head
2025-12-04T08:53:09.5558953Z  * [new branch]                gh/williamwen42/335/orig -> origin/gh/williamwen42/335/orig
2025-12-04T08:53:09.5559147Z  * [new branch]                gh/williamwen42/336/base -> origin/gh/williamwen42/336/base
2025-12-04T08:53:09.5559339Z  * [new branch]                gh/williamwen42/336/head -> origin/gh/williamwen42/336/head
2025-12-04T08:53:09.5559532Z  * [new branch]                gh/williamwen42/336/orig -> origin/gh/williamwen42/336/orig
2025-12-04T08:53:09.5559723Z  * [new branch]                gh/williamwen42/337/base -> origin/gh/williamwen42/337/base
2025-12-04T08:53:09.5559914Z  * [new branch]                gh/williamwen42/337/head -> origin/gh/williamwen42/337/head
2025-12-04T08:53:09.5560106Z  * [new branch]                gh/williamwen42/337/orig -> origin/gh/williamwen42/337/orig
2025-12-04T08:53:09.5560296Z  * [new branch]                gh/williamwen42/338/base -> origin/gh/williamwen42/338/base
2025-12-04T08:53:09.5560553Z  * [new branch]                gh/williamwen42/338/head -> origin/gh/williamwen42/338/head
2025-12-04T08:53:09.5560804Z  * [new branch]                gh/williamwen42/338/orig -> origin/gh/williamwen42/338/orig
2025-12-04T08:53:09.5560998Z  * [new branch]                gh/williamwen42/339/base -> origin/gh/williamwen42/339/base
2025-12-04T08:53:09.5561189Z  * [new branch]                gh/williamwen42/339/head -> origin/gh/williamwen42/339/head
2025-12-04T08:53:09.5561383Z  * [new branch]                gh/williamwen42/339/orig -> origin/gh/williamwen42/339/orig
2025-12-04T08:53:09.5561575Z  * [new branch]                gh/williamwen42/340/base -> origin/gh/williamwen42/340/base
2025-12-04T08:53:09.5561765Z  * [new branch]                gh/williamwen42/340/head -> origin/gh/williamwen42/340/head
2025-12-04T08:53:09.5561958Z  * [new branch]                gh/williamwen42/340/orig -> origin/gh/williamwen42/340/orig
2025-12-04T08:53:09.5562153Z  * [new branch]                gh/williamwen42/341/base -> origin/gh/williamwen42/341/base
2025-12-04T08:53:09.5562344Z  * [new branch]                gh/williamwen42/341/head -> origin/gh/williamwen42/341/head
2025-12-04T08:53:09.5562544Z  * [new branch]                gh/williamwen42/341/orig -> origin/gh/williamwen42/341/orig
2025-12-04T08:53:09.5562736Z  * [new branch]                gh/williamwen42/342/base -> origin/gh/williamwen42/342/base
2025-12-04T08:53:09.5562926Z  * [new branch]                gh/williamwen42/342/head -> origin/gh/williamwen42/342/head
2025-12-04T08:53:09.5563119Z  * [new branch]                gh/williamwen42/342/orig -> origin/gh/williamwen42/342/orig
2025-12-04T08:53:09.5563312Z  * [new branch]                gh/williamwen42/343/base -> origin/gh/williamwen42/343/base
2025-12-04T08:53:09.5563503Z  * [new branch]                gh/williamwen42/343/head -> origin/gh/williamwen42/343/head
2025-12-04T08:53:09.5563694Z  * [new branch]                gh/williamwen42/343/orig -> origin/gh/williamwen42/343/orig
2025-12-04T08:53:09.5563889Z  * [new branch]                gh/williamwen42/344/base -> origin/gh/williamwen42/344/base
2025-12-04T08:53:09.5564086Z  * [new branch]                gh/williamwen42/344/head -> origin/gh/williamwen42/344/head
2025-12-04T08:53:09.5564280Z  * [new branch]                gh/williamwen42/344/orig -> origin/gh/williamwen42/344/orig
2025-12-04T08:53:09.5564474Z  * [new branch]                gh/williamwen42/345/base -> origin/gh/williamwen42/345/base
2025-12-04T08:53:09.5564667Z  * [new branch]                gh/williamwen42/345/head -> origin/gh/williamwen42/345/head
2025-12-04T08:53:09.5564858Z  * [new branch]                gh/williamwen42/345/orig -> origin/gh/williamwen42/345/orig
2025-12-04T08:53:09.5565051Z  * [new branch]                gh/williamwen42/346/base -> origin/gh/williamwen42/346/base
2025-12-04T08:53:09.5565242Z  * [new branch]                gh/williamwen42/346/head -> origin/gh/williamwen42/346/head
2025-12-04T08:53:09.5565437Z  * [new branch]                gh/williamwen42/346/orig -> origin/gh/williamwen42/346/orig
2025-12-04T08:53:09.5565628Z  * [new branch]                gh/williamwen42/347/base -> origin/gh/williamwen42/347/base
2025-12-04T08:53:09.5565823Z  * [new branch]                gh/williamwen42/347/head -> origin/gh/williamwen42/347/head
2025-12-04T08:53:09.5566016Z  * [new branch]                gh/williamwen42/347/orig -> origin/gh/williamwen42/347/orig
2025-12-04T08:53:09.5566205Z  * [new branch]                gh/williamwen42/348/base -> origin/gh/williamwen42/348/base
2025-12-04T08:53:09.5566398Z  * [new branch]                gh/williamwen42/348/head -> origin/gh/williamwen42/348/head
2025-12-04T08:53:09.5566589Z  * [new branch]                gh/williamwen42/348/orig -> origin/gh/williamwen42/348/orig
2025-12-04T08:53:09.5566778Z  * [new branch]                gh/williamwen42/349/base -> origin/gh/williamwen42/349/base
2025-12-04T08:53:09.5566971Z  * [new branch]                gh/williamwen42/349/head -> origin/gh/williamwen42/349/head
2025-12-04T08:53:09.5567213Z  * [new branch]                gh/williamwen42/349/orig -> origin/gh/williamwen42/349/orig
2025-12-04T08:53:09.5567433Z  * [new branch]                gh/williamwen42/350/base -> origin/gh/williamwen42/350/base
2025-12-04T08:53:09.5567625Z  * [new branch]                gh/williamwen42/350/head -> origin/gh/williamwen42/350/head
2025-12-04T08:53:09.5567818Z  * [new branch]                gh/williamwen42/350/orig -> origin/gh/williamwen42/350/orig
2025-12-04T08:53:09.5568010Z  * [new branch]                gh/williamwen42/351/base -> origin/gh/williamwen42/351/base
2025-12-04T08:53:09.5568202Z  * [new branch]                gh/williamwen42/351/head -> origin/gh/williamwen42/351/head
2025-12-04T08:53:09.5568396Z  * [new branch]                gh/williamwen42/351/orig -> origin/gh/williamwen42/351/orig
2025-12-04T08:53:09.5568586Z  * [new branch]                gh/williamwen42/352/base -> origin/gh/williamwen42/352/base
2025-12-04T08:53:09.5568782Z  * [new branch]                gh/williamwen42/352/head -> origin/gh/williamwen42/352/head
2025-12-04T08:53:09.5568973Z  * [new branch]                gh/williamwen42/352/orig -> origin/gh/williamwen42/352/orig
2025-12-04T08:53:09.5569168Z  * [new branch]                gh/williamwen42/353/base -> origin/gh/williamwen42/353/base
2025-12-04T08:53:09.5569360Z  * [new branch]                gh/williamwen42/353/head -> origin/gh/williamwen42/353/head
2025-12-04T08:53:09.5569552Z  * [new branch]                gh/williamwen42/353/orig -> origin/gh/williamwen42/353/orig
2025-12-04T08:53:09.5569746Z  * [new branch]                gh/williamwen42/354/base -> origin/gh/williamwen42/354/base
2025-12-04T08:53:09.5569937Z  * [new branch]                gh/williamwen42/354/head -> origin/gh/williamwen42/354/head
2025-12-04T08:53:09.5570129Z  * [new branch]                gh/williamwen42/354/orig -> origin/gh/williamwen42/354/orig
2025-12-04T08:53:09.5570320Z  * [new branch]                gh/williamwen42/355/base -> origin/gh/williamwen42/355/base
2025-12-04T08:53:09.5570540Z  * [new branch]                gh/williamwen42/355/head -> origin/gh/williamwen42/355/head
2025-12-04T08:53:09.5570739Z  * [new branch]                gh/williamwen42/355/orig -> origin/gh/williamwen42/355/orig
2025-12-04T08:53:09.5570932Z  * [new branch]                gh/williamwen42/356/base -> origin/gh/williamwen42/356/base
2025-12-04T08:53:09.5571123Z  * [new branch]                gh/williamwen42/356/head -> origin/gh/williamwen42/356/head
2025-12-04T08:53:09.5571314Z  * [new branch]                gh/williamwen42/356/orig -> origin/gh/williamwen42/356/orig
2025-12-04T08:53:09.5571505Z  * [new branch]                gh/williamwen42/357/base -> origin/gh/williamwen42/357/base
2025-12-04T08:53:09.5571696Z  * [new branch]                gh/williamwen42/357/head -> origin/gh/williamwen42/357/head
2025-12-04T08:53:09.5571886Z  * [new branch]                gh/williamwen42/357/orig -> origin/gh/williamwen42/357/orig
2025-12-04T08:53:09.5572081Z  * [new branch]                gh/williamwen42/358/base -> origin/gh/williamwen42/358/base
2025-12-04T08:53:09.5572273Z  * [new branch]                gh/williamwen42/358/head -> origin/gh/williamwen42/358/head
2025-12-04T08:53:09.5572467Z  * [new branch]                gh/williamwen42/358/orig -> origin/gh/williamwen42/358/orig
2025-12-04T08:53:09.5572652Z  * [new branch]                gh/xmfan/169/base       -> origin/gh/xmfan/169/base
2025-12-04T08:53:09.5572827Z  * [new branch]                gh/xmfan/169/head       -> origin/gh/xmfan/169/head
2025-12-04T08:53:09.5572998Z  * [new branch]                gh/xmfan/170/base       -> origin/gh/xmfan/170/base
2025-12-04T08:53:09.5573169Z  * [new branch]                gh/xmfan/170/head       -> origin/gh/xmfan/170/head
2025-12-04T08:53:09.5573341Z  * [new branch]                gh/xmfan/274/base       -> origin/gh/xmfan/274/base
2025-12-04T08:53:09.5573507Z  * [new branch]                gh/xmfan/274/head       -> origin/gh/xmfan/274/head
2025-12-04T08:53:09.5573729Z  * [new branch]                gh/xmfan/274/orig       -> origin/gh/xmfan/274/orig
2025-12-04T08:53:09.5573897Z  * [new branch]                gh/xmfan/277/base       -> origin/gh/xmfan/277/base
2025-12-04T08:53:09.5574117Z  * [new branch]                gh/xmfan/277/head       -> origin/gh/xmfan/277/head
2025-12-04T08:53:09.5574286Z  * [new branch]                gh/xmfan/277/orig       -> origin/gh/xmfan/277/orig
2025-12-04T08:53:09.5574453Z  * [new branch]                gh/xmfan/301/base       -> origin/gh/xmfan/301/base
2025-12-04T08:53:09.5574620Z  * [new branch]                gh/xmfan/301/head       -> origin/gh/xmfan/301/head
2025-12-04T08:53:09.5574790Z  * [new branch]                gh/xmfan/301/orig       -> origin/gh/xmfan/301/orig
2025-12-04T08:53:09.5574960Z  * [new branch]                gh/xmfan/304/base       -> origin/gh/xmfan/304/base
2025-12-04T08:53:09.5575126Z  * [new branch]                gh/xmfan/304/head       -> origin/gh/xmfan/304/head
2025-12-04T08:53:09.5575295Z  * [new branch]                gh/xmfan/304/orig       -> origin/gh/xmfan/304/orig
2025-12-04T08:53:09.5575468Z  * [new branch]                gh/xmfan/309/base       -> origin/gh/xmfan/309/base
2025-12-04T08:53:09.5575648Z  * [new branch]                gh/xmfan/309/head       -> origin/gh/xmfan/309/head
2025-12-04T08:53:09.5575816Z  * [new branch]                gh/xmfan/309/orig       -> origin/gh/xmfan/309/orig
2025-12-04T08:53:09.5575985Z  * [new branch]                gh/xmfan/310/base       -> origin/gh/xmfan/310/base
2025-12-04T08:53:09.5576153Z  * [new branch]                gh/xmfan/310/head       -> origin/gh/xmfan/310/head
2025-12-04T08:53:09.5576323Z  * [new branch]                gh/xmfan/310/orig       -> origin/gh/xmfan/310/orig
2025-12-04T08:53:09.5576491Z  * [new branch]                gh/xmfan/311/base       -> origin/gh/xmfan/311/base
2025-12-04T08:53:09.5576659Z  * [new branch]                gh/xmfan/311/head       -> origin/gh/xmfan/311/head
2025-12-04T08:53:09.5576827Z  * [new branch]                gh/xmfan/311/orig       -> origin/gh/xmfan/311/orig
2025-12-04T08:53:09.5576996Z  * [new branch]                gh/xmfan/312/base       -> origin/gh/xmfan/312/base
2025-12-04T08:53:09.5577169Z  * [new branch]                gh/xmfan/312/head       -> origin/gh/xmfan/312/head
2025-12-04T08:53:09.5577338Z  * [new branch]                gh/xmfan/312/orig       -> origin/gh/xmfan/312/orig
2025-12-04T08:53:09.5577505Z  * [new branch]                gh/xmfan/313/base       -> origin/gh/xmfan/313/base
2025-12-04T08:53:09.5577673Z  * [new branch]                gh/xmfan/313/head       -> origin/gh/xmfan/313/head
2025-12-04T08:53:09.5577843Z  * [new branch]                gh/xmfan/313/orig       -> origin/gh/xmfan/313/orig
2025-12-04T08:53:09.5578022Z  * [new branch]                gh/xuanzhang816/27/base -> origin/gh/xuanzhang816/27/base
2025-12-04T08:53:09.5578214Z  * [new branch]                gh/xuanzhang816/27/head -> origin/gh/xuanzhang816/27/head
2025-12-04T08:53:09.5578402Z  * [new branch]                gh/xuanzhang816/27/orig -> origin/gh/xuanzhang816/27/orig
2025-12-04T08:53:09.5578595Z  * [new branch]                gh/xuanzhang816/32/base -> origin/gh/xuanzhang816/32/base
2025-12-04T08:53:09.5578785Z  * [new branch]                gh/xuanzhang816/32/head -> origin/gh/xuanzhang816/32/head
2025-12-04T08:53:09.5578970Z  * [new branch]                gh/xuanzhang816/32/orig -> origin/gh/xuanzhang816/32/orig
2025-12-04T08:53:09.5579157Z  * [new branch]                gh/xuanzhang816/33/base -> origin/gh/xuanzhang816/33/base
2025-12-04T08:53:09.5579345Z  * [new branch]                gh/xuanzhang816/33/head -> origin/gh/xuanzhang816/33/head
2025-12-04T08:53:09.5579530Z  * [new branch]                gh/xuanzhang816/33/orig -> origin/gh/xuanzhang816/33/orig
2025-12-04T08:53:09.5579719Z  * [new branch]                gh/xuanzhang816/34/base -> origin/gh/xuanzhang816/34/base
2025-12-04T08:53:09.5579905Z  * [new branch]                gh/xuanzhang816/34/head -> origin/gh/xuanzhang816/34/head
2025-12-04T08:53:09.5580128Z  * [new branch]                gh/xuanzhang816/34/orig -> origin/gh/xuanzhang816/34/orig
2025-12-04T08:53:09.5580316Z  * [new branch]                gh/xuanzhang816/35/base -> origin/gh/xuanzhang816/35/base
2025-12-04T08:53:09.5580577Z  * [new branch]                gh/xuanzhang816/35/head -> origin/gh/xuanzhang816/35/head
2025-12-04T08:53:09.5580763Z  * [new branch]                gh/xuanzhang816/35/orig -> origin/gh/xuanzhang816/35/orig
2025-12-04T08:53:09.5580949Z  * [new branch]                gh/yanbing-j/11/base    -> origin/gh/yanbing-j/11/base
2025-12-04T08:53:09.5581129Z  * [new branch]                gh/yanbing-j/11/head    -> origin/gh/yanbing-j/11/head
2025-12-04T08:53:09.5581309Z  * [new branch]                gh/yanbing-j/11/orig    -> origin/gh/yanbing-j/11/orig
2025-12-04T08:53:09.5581487Z  * [new branch]                gh/yanbing-j/12/base    -> origin/gh/yanbing-j/12/base
2025-12-04T08:53:09.5581664Z  * [new branch]                gh/yanbing-j/12/head    -> origin/gh/yanbing-j/12/head
2025-12-04T08:53:09.5581842Z  * [new branch]                gh/yanbing-j/12/orig    -> origin/gh/yanbing-j/12/orig
2025-12-04T08:53:09.5582019Z  * [new branch]                gh/yanbing-j/13/base    -> origin/gh/yanbing-j/13/base
2025-12-04T08:53:09.5582200Z  * [new branch]                gh/yanbing-j/13/head    -> origin/gh/yanbing-j/13/head
2025-12-04T08:53:09.5582376Z  * [new branch]                gh/yanbing-j/13/orig    -> origin/gh/yanbing-j/13/orig
2025-12-04T08:53:09.5582553Z  * [new branch]                gh/yanbing-j/14/base    -> origin/gh/yanbing-j/14/base
2025-12-04T08:53:09.5582728Z  * [new branch]                gh/yanbing-j/14/head    -> origin/gh/yanbing-j/14/head
2025-12-04T08:53:09.5582902Z  * [new branch]                gh/yanbing-j/14/orig    -> origin/gh/yanbing-j/14/orig
2025-12-04T08:53:09.5583077Z  * [new branch]                gh/yanbing-j/15/base    -> origin/gh/yanbing-j/15/base
2025-12-04T08:53:09.5583253Z  * [new branch]                gh/yanbing-j/15/head    -> origin/gh/yanbing-j/15/head
2025-12-04T08:53:09.5583428Z  * [new branch]                gh/yanbing-j/15/orig    -> origin/gh/yanbing-j/15/orig
2025-12-04T08:53:09.5583604Z  * [new branch]                gh/yanbing-j/18/base    -> origin/gh/yanbing-j/18/base
2025-12-04T08:53:09.5583783Z  * [new branch]                gh/yanbing-j/18/head    -> origin/gh/yanbing-j/18/head
2025-12-04T08:53:09.5583959Z  * [new branch]                gh/yanbing-j/18/orig    -> origin/gh/yanbing-j/18/orig
2025-12-04T08:53:09.5584134Z  * [new branch]                gh/yanbing-j/19/base    -> origin/gh/yanbing-j/19/base
2025-12-04T08:53:09.5584307Z  * [new branch]                gh/yanbing-j/19/head    -> origin/gh/yanbing-j/19/head
2025-12-04T08:53:09.5584482Z  * [new branch]                gh/yanbing-j/19/orig    -> origin/gh/yanbing-j/19/orig
2025-12-04T08:53:09.5584658Z  * [new branch]                gh/yanbing-j/20/base    -> origin/gh/yanbing-j/20/base
2025-12-04T08:53:09.5584831Z  * [new branch]                gh/yanbing-j/20/head    -> origin/gh/yanbing-j/20/head
2025-12-04T08:53:09.5585007Z  * [new branch]                gh/yanbing-j/20/orig    -> origin/gh/yanbing-j/20/orig
2025-12-04T08:53:09.5585185Z  * [new branch]                gh/yanbing-j/21/base    -> origin/gh/yanbing-j/21/base
2025-12-04T08:53:09.5585360Z  * [new branch]                gh/yanbing-j/21/head    -> origin/gh/yanbing-j/21/head
2025-12-04T08:53:09.5585534Z  * [new branch]                gh/yanbing-j/22/base    -> origin/gh/yanbing-j/22/base
2025-12-04T08:53:09.5585713Z  * [new branch]                gh/yanbing-j/22/head    -> origin/gh/yanbing-j/22/head
2025-12-04T08:53:09.5585887Z  * [new branch]                gh/yanbing-j/22/orig    -> origin/gh/yanbing-j/22/orig
2025-12-04T08:53:09.5586062Z  * [new branch]                gh/yanbing-j/23/base    -> origin/gh/yanbing-j/23/base
2025-12-04T08:53:09.5586239Z  * [new branch]                gh/yanbing-j/23/head    -> origin/gh/yanbing-j/23/head
2025-12-04T08:53:09.5586414Z  * [new branch]                gh/yanbing-j/23/orig    -> origin/gh/yanbing-j/23/orig
2025-12-04T08:53:09.5586682Z  * [new branch]                gh/yanbing-j/24/base    -> origin/gh/yanbing-j/24/base
2025-12-04T08:53:09.5586891Z  * [new branch]                gh/yanbing-j/24/head    -> origin/gh/yanbing-j/24/head
2025-12-04T08:53:09.5587067Z  * [new branch]                gh/yanbing-j/24/orig    -> origin/gh/yanbing-j/24/orig
2025-12-04T08:53:09.5587243Z  * [new branch]                gh/yanbing-j/25/base    -> origin/gh/yanbing-j/25/base
2025-12-04T08:53:09.5587420Z  * [new branch]                gh/yanbing-j/25/head    -> origin/gh/yanbing-j/25/head
2025-12-04T08:53:09.5587595Z  * [new branch]                gh/yanbing-j/25/orig    -> origin/gh/yanbing-j/25/orig
2025-12-04T08:53:09.5587770Z  * [new branch]                gh/yanbing-j/26/base    -> origin/gh/yanbing-j/26/base
2025-12-04T08:53:09.5587943Z  * [new branch]                gh/yanbing-j/26/head    -> origin/gh/yanbing-j/26/head
2025-12-04T08:53:09.5588119Z  * [new branch]                gh/yanbing-j/26/orig    -> origin/gh/yanbing-j/26/orig
2025-12-04T08:53:09.5588309Z  * [new branch]                gh/yang-yu-hang/1/base  -> origin/gh/yang-yu-hang/1/base
2025-12-04T08:53:09.5588499Z  * [new branch]                gh/yang-yu-hang/1/head  -> origin/gh/yang-yu-hang/1/head
2025-12-04T08:53:09.5588684Z  * [new branch]                gh/yang-yu-hang/1/orig  -> origin/gh/yang-yu-hang/1/orig
2025-12-04T08:53:09.5588868Z  * [new branch]                gh/yang-yu-hang/2/base  -> origin/gh/yang-yu-hang/2/base
2025-12-04T08:53:09.5589051Z  * [new branch]                gh/yang-yu-hang/2/head  -> origin/gh/yang-yu-hang/2/head
2025-12-04T08:53:09.5589233Z  * [new branch]                gh/yang-yu-hang/2/orig  -> origin/gh/yang-yu-hang/2/orig
2025-12-04T08:53:09.5589417Z  * [new branch]                gh/yang-yu-hang/3/base  -> origin/gh/yang-yu-hang/3/base
2025-12-04T08:53:09.5589597Z  * [new branch]                gh/yang-yu-hang/3/head  -> origin/gh/yang-yu-hang/3/head
2025-12-04T08:53:09.5589778Z  * [new branch]                gh/yang-yu-hang/3/orig  -> origin/gh/yang-yu-hang/3/orig
2025-12-04T08:53:09.5589963Z  * [new branch]                gh/yangw-dev/12/base    -> origin/gh/yangw-dev/12/base
2025-12-04T08:53:09.5590145Z  * [new branch]                gh/yangw-dev/12/head    -> origin/gh/yangw-dev/12/head
2025-12-04T08:53:09.5590324Z  * [new branch]                gh/yangw-dev/12/orig    -> origin/gh/yangw-dev/12/orig
2025-12-04T08:53:09.5590535Z  * [new branch]                gh/yangw-dev/13/base    -> origin/gh/yangw-dev/13/base
2025-12-04T08:53:09.5590709Z  * [new branch]                gh/yangw-dev/13/head    -> origin/gh/yangw-dev/13/head
2025-12-04T08:53:09.5590884Z  * [new branch]                gh/yangw-dev/13/orig    -> origin/gh/yangw-dev/13/orig
2025-12-04T08:53:09.5591059Z  * [new branch]                gh/yangw-dev/14/base    -> origin/gh/yangw-dev/14/base
2025-12-04T08:53:09.5591233Z  * [new branch]                gh/yangw-dev/14/head    -> origin/gh/yangw-dev/14/head
2025-12-04T08:53:09.5591413Z  * [new branch]                gh/yangw-dev/14/orig    -> origin/gh/yangw-dev/14/orig
2025-12-04T08:53:09.5591588Z  * [new branch]                gh/yangw-dev/15/base    -> origin/gh/yangw-dev/15/base
2025-12-04T08:53:09.5591768Z  * [new branch]                gh/yangw-dev/15/head    -> origin/gh/yangw-dev/15/head
2025-12-04T08:53:09.5591944Z  * [new branch]                gh/yangw-dev/15/orig    -> origin/gh/yangw-dev/15/orig
2025-12-04T08:53:09.5592117Z  * [new branch]                gh/yangw-dev/19/base    -> origin/gh/yangw-dev/19/base
2025-12-04T08:53:09.5592292Z  * [new branch]                gh/yangw-dev/19/head    -> origin/gh/yangw-dev/19/head
2025-12-04T08:53:09.5592467Z  * [new branch]                gh/yangw-dev/19/orig    -> origin/gh/yangw-dev/19/orig
2025-12-04T08:53:09.5592640Z  * [new branch]                gh/yangw-dev/26/base    -> origin/gh/yangw-dev/26/base
2025-12-04T08:53:09.5592815Z  * [new branch]                gh/yangw-dev/26/head    -> origin/gh/yangw-dev/26/head
2025-12-04T08:53:09.5593046Z  * [new branch]                gh/yangw-dev/26/orig    -> origin/gh/yangw-dev/26/orig
2025-12-04T08:53:09.5593221Z  * [new branch]                gh/yangw-dev/27/base    -> origin/gh/yangw-dev/27/base
2025-12-04T08:53:09.5593459Z  * [new branch]                gh/yangw-dev/27/head    -> origin/gh/yangw-dev/27/head
2025-12-04T08:53:09.5593637Z  * [new branch]                gh/yangw-dev/27/orig    -> origin/gh/yangw-dev/27/orig
2025-12-04T08:53:09.5593811Z  * [new branch]                gh/ydwu4/292/base       -> origin/gh/ydwu4/292/base
2025-12-04T08:53:09.5593984Z  * [new branch]                gh/ydwu4/292/head       -> origin/gh/ydwu4/292/head
2025-12-04T08:53:09.5594154Z  * [new branch]                gh/ydwu4/292/orig       -> origin/gh/ydwu4/292/orig
2025-12-04T08:53:09.5594323Z  * [new branch]                gh/ydwu4/294/base       -> origin/gh/ydwu4/294/base
2025-12-04T08:53:09.5594492Z  * [new branch]                gh/ydwu4/294/head       -> origin/gh/ydwu4/294/head
2025-12-04T08:53:09.5594664Z  * [new branch]                gh/ydwu4/294/orig       -> origin/gh/ydwu4/294/orig
2025-12-04T08:53:09.5594831Z  * [new branch]                gh/ydwu4/295/base       -> origin/gh/ydwu4/295/base
2025-12-04T08:53:09.5595003Z  * [new branch]                gh/ydwu4/295/head       -> origin/gh/ydwu4/295/head
2025-12-04T08:53:09.5595172Z  * [new branch]                gh/ydwu4/295/orig       -> origin/gh/ydwu4/295/orig
2025-12-04T08:53:09.5595339Z  * [new branch]                gh/ydwu4/296/base       -> origin/gh/ydwu4/296/base
2025-12-04T08:53:09.5595506Z  * [new branch]                gh/ydwu4/296/head       -> origin/gh/ydwu4/296/head
2025-12-04T08:53:09.5595673Z  * [new branch]                gh/ydwu4/296/orig       -> origin/gh/ydwu4/296/orig
2025-12-04T08:53:09.5595842Z  * [new branch]                gh/ydwu4/306/base       -> origin/gh/ydwu4/306/base
2025-12-04T08:53:09.5596010Z  * [new branch]                gh/ydwu4/306/head       -> origin/gh/ydwu4/306/head
2025-12-04T08:53:09.5596176Z  * [new branch]                gh/ydwu4/306/orig       -> origin/gh/ydwu4/306/orig
2025-12-04T08:53:09.5596347Z  * [new branch]                gh/ydwu4/312/base       -> origin/gh/ydwu4/312/base
2025-12-04T08:53:09.5596518Z  * [new branch]                gh/ydwu4/312/head       -> origin/gh/ydwu4/312/head
2025-12-04T08:53:09.5596685Z  * [new branch]                gh/ydwu4/312/orig       -> origin/gh/ydwu4/312/orig
2025-12-04T08:53:09.5596853Z  * [new branch]                gh/ydwu4/322/base       -> origin/gh/ydwu4/322/base
2025-12-04T08:53:09.5597021Z  * [new branch]                gh/ydwu4/322/head       -> origin/gh/ydwu4/322/head
2025-12-04T08:53:09.5597188Z  * [new branch]                gh/ydwu4/322/orig       -> origin/gh/ydwu4/322/orig
2025-12-04T08:53:09.5597357Z  * [new branch]                gh/ydwu4/327/base       -> origin/gh/ydwu4/327/base
2025-12-04T08:53:09.5597524Z  * [new branch]                gh/ydwu4/327/head       -> origin/gh/ydwu4/327/head
2025-12-04T08:53:09.5597691Z  * [new branch]                gh/ydwu4/327/orig       -> origin/gh/ydwu4/327/orig
2025-12-04T08:53:09.5597860Z  * [new branch]                gh/ydwu4/328/base       -> origin/gh/ydwu4/328/base
2025-12-04T08:53:09.5598033Z  * [new branch]                gh/ydwu4/328/head       -> origin/gh/ydwu4/328/head
2025-12-04T08:53:09.5598201Z  * [new branch]                gh/ydwu4/328/orig       -> origin/gh/ydwu4/328/orig
2025-12-04T08:53:09.5598369Z  * [new branch]                gh/ydwu4/329/base       -> origin/gh/ydwu4/329/base
2025-12-04T08:53:09.5598538Z  * [new branch]                gh/ydwu4/329/head       -> origin/gh/ydwu4/329/head
2025-12-04T08:53:09.5598705Z  * [new branch]                gh/ydwu4/329/orig       -> origin/gh/ydwu4/329/orig
2025-12-04T08:53:09.5598873Z  * [new branch]                gh/ydwu4/330/base       -> origin/gh/ydwu4/330/base
2025-12-04T08:53:09.5599041Z  * [new branch]                gh/ydwu4/330/head       -> origin/gh/ydwu4/330/head
2025-12-04T08:53:09.5599209Z  * [new branch]                gh/ydwu4/330/orig       -> origin/gh/ydwu4/330/orig
2025-12-04T08:53:09.5599416Z  * [new branch]                gh/ydwu4/331/base       -> origin/gh/ydwu4/331/base
2025-12-04T08:53:09.5599609Z  * [new branch]                gh/ydwu4/331/head       -> origin/gh/ydwu4/331/head
2025-12-04T08:53:09.5599777Z  * [new branch]                gh/ydwu4/331/orig       -> origin/gh/ydwu4/331/orig
2025-12-04T08:53:09.5599945Z  * [new branch]                gh/ydwu4/332/base       -> origin/gh/ydwu4/332/base
2025-12-04T08:53:09.5600113Z  * [new branch]                gh/ydwu4/332/head       -> origin/gh/ydwu4/332/head
2025-12-04T08:53:09.5600284Z  * [new branch]                gh/ydwu4/332/orig       -> origin/gh/ydwu4/332/orig
2025-12-04T08:53:09.5600487Z  * [new branch]                gh/ydwu4/333/base       -> origin/gh/ydwu4/333/base
2025-12-04T08:53:09.5600656Z  * [new branch]                gh/ydwu4/333/head       -> origin/gh/ydwu4/333/head
2025-12-04T08:53:09.5600825Z  * [new branch]                gh/ydwu4/333/orig       -> origin/gh/ydwu4/333/orig
2025-12-04T08:53:09.5600996Z  * [new branch]                gh/ydwu4/334/base       -> origin/gh/ydwu4/334/base
2025-12-04T08:53:09.5601166Z  * [new branch]                gh/ydwu4/334/head       -> origin/gh/ydwu4/334/head
2025-12-04T08:53:09.5601335Z  * [new branch]                gh/ydwu4/334/orig       -> origin/gh/ydwu4/334/orig
2025-12-04T08:53:09.5601503Z  * [new branch]                gh/ydwu4/335/base       -> origin/gh/ydwu4/335/base
2025-12-04T08:53:09.5601670Z  * [new branch]                gh/ydwu4/335/head       -> origin/gh/ydwu4/335/head
2025-12-04T08:53:09.5601838Z  * [new branch]                gh/ydwu4/335/orig       -> origin/gh/ydwu4/335/orig
2025-12-04T08:53:09.5602007Z  * [new branch]                gh/ydwu4/337/base       -> origin/gh/ydwu4/337/base
2025-12-04T08:53:09.5602173Z  * [new branch]                gh/ydwu4/337/head       -> origin/gh/ydwu4/337/head
2025-12-04T08:53:09.5602340Z  * [new branch]                gh/ydwu4/337/orig       -> origin/gh/ydwu4/337/orig
2025-12-04T08:53:09.5602508Z  * [new branch]                gh/ydwu4/339/base       -> origin/gh/ydwu4/339/base
2025-12-04T08:53:09.5602676Z  * [new branch]                gh/ydwu4/339/head       -> origin/gh/ydwu4/339/head
2025-12-04T08:53:09.5602847Z  * [new branch]                gh/ydwu4/339/orig       -> origin/gh/ydwu4/339/orig
2025-12-04T08:53:09.5603013Z  * [new branch]                gh/yf225/133/base       -> origin/gh/yf225/133/base
2025-12-04T08:53:09.5603179Z  * [new branch]                gh/yf225/133/head       -> origin/gh/yf225/133/head
2025-12-04T08:53:09.5603347Z  * [new branch]                gh/yf225/93/base        -> origin/gh/yf225/93/base
2025-12-04T08:53:09.5603512Z  * [new branch]                gh/yf225/93/head        -> origin/gh/yf225/93/head
2025-12-04T08:53:09.5603687Z  * [new branch]                gh/yifuwang/152/base    -> origin/gh/yifuwang/152/base
2025-12-04T08:53:09.5603869Z  * [new branch]                gh/yifuwang/152/head    -> origin/gh/yifuwang/152/head
2025-12-04T08:53:09.5604053Z  * [new branch]                gh/yifuwang/152/orig    -> origin/gh/yifuwang/152/orig
2025-12-04T08:53:09.5604232Z  * [new branch]                gh/yifuwang/195/base    -> origin/gh/yifuwang/195/base
2025-12-04T08:53:09.5604414Z  * [new branch]                gh/yifuwang/195/head    -> origin/gh/yifuwang/195/head
2025-12-04T08:53:09.5604592Z  * [new branch]                gh/yifuwang/195/orig    -> origin/gh/yifuwang/195/orig
2025-12-04T08:53:09.5604773Z  * [new branch]                gh/yiming0416/1/base    -> origin/gh/yiming0416/1/base
2025-12-04T08:53:09.5604950Z  * [new branch]                gh/yiming0416/1/head    -> origin/gh/yiming0416/1/head
2025-12-04T08:53:09.5605126Z  * [new branch]                gh/yiming0416/2/base    -> origin/gh/yiming0416/2/base
2025-12-04T08:53:09.5605302Z  * [new branch]                gh/yiming0416/2/head    -> origin/gh/yiming0416/2/head
2025-12-04T08:53:09.5605483Z  * [new branch]                gh/yushangdi/1/base     -> origin/gh/yushangdi/1/base
2025-12-04T08:53:09.5605721Z  * [new branch]                gh/yushangdi/1/head     -> origin/gh/yushangdi/1/head
2025-12-04T08:53:09.5605902Z  * [new branch]                gh/yushangdi/10/base    -> origin/gh/yushangdi/10/base
2025-12-04T08:53:09.5606127Z  * [new branch]                gh/yushangdi/10/head    -> origin/gh/yushangdi/10/head
2025-12-04T08:53:09.5606308Z  * [new branch]                gh/yushangdi/10/orig    -> origin/gh/yushangdi/10/orig
2025-12-04T08:53:09.5606487Z  * [new branch]                gh/yushangdi/11/base    -> origin/gh/yushangdi/11/base
2025-12-04T08:53:09.5606667Z  * [new branch]                gh/yushangdi/11/head    -> origin/gh/yushangdi/11/head
2025-12-04T08:53:09.5606848Z  * [new branch]                gh/yushangdi/11/orig    -> origin/gh/yushangdi/11/orig
2025-12-04T08:53:09.5607028Z  * [new branch]                gh/yushangdi/2/base     -> origin/gh/yushangdi/2/base
2025-12-04T08:53:09.5607205Z  * [new branch]                gh/yushangdi/2/head     -> origin/gh/yushangdi/2/head
2025-12-04T08:53:09.5607389Z  * [new branch]                gh/yushangdi/7/base     -> origin/gh/yushangdi/7/base
2025-12-04T08:53:09.5607566Z  * [new branch]                gh/yushangdi/7/head     -> origin/gh/yushangdi/7/head
2025-12-04T08:53:09.5607743Z  * [new branch]                gh/yushangdi/7/orig     -> origin/gh/yushangdi/7/orig
2025-12-04T08:53:09.5607919Z  * [new branch]                gh/yushangdi/8/base     -> origin/gh/yushangdi/8/base
2025-12-04T08:53:09.5608096Z  * [new branch]                gh/yushangdi/8/head     -> origin/gh/yushangdi/8/head
2025-12-04T08:53:09.5608271Z  * [new branch]                gh/yushangdi/8/orig     -> origin/gh/yushangdi/8/orig
2025-12-04T08:53:09.5608447Z  * [new branch]                gh/yushangdi/9/base     -> origin/gh/yushangdi/9/base
2025-12-04T08:53:09.5608625Z  * [new branch]                gh/yushangdi/9/head     -> origin/gh/yushangdi/9/head
2025-12-04T08:53:09.5608802Z  * [new branch]                gh/yushangdi/9/orig     -> origin/gh/yushangdi/9/orig
2025-12-04T08:53:09.5608981Z  * [new branch]                gh/zklaus/19/base       -> origin/gh/zklaus/19/base
2025-12-04T08:53:09.5609152Z  * [new branch]                gh/zklaus/19/head       -> origin/gh/zklaus/19/head
2025-12-04T08:53:09.5609325Z  * [new branch]                gh/zklaus/19/orig       -> origin/gh/zklaus/19/orig
2025-12-04T08:53:09.5609495Z  * [new branch]                gh/zklaus/20/base       -> origin/gh/zklaus/20/base
2025-12-04T08:53:09.5609663Z  * [new branch]                gh/zklaus/20/head       -> origin/gh/zklaus/20/head
2025-12-04T08:53:09.5609831Z  * [new branch]                gh/zklaus/20/orig       -> origin/gh/zklaus/20/orig
2025-12-04T08:53:09.5609999Z  * [new branch]                gh/zklaus/21/base       -> origin/gh/zklaus/21/base
2025-12-04T08:53:09.5610167Z  * [new branch]                gh/zklaus/21/head       -> origin/gh/zklaus/21/head
2025-12-04T08:53:09.5610336Z  * [new branch]                gh/zklaus/21/orig       -> origin/gh/zklaus/21/orig
2025-12-04T08:53:09.5610549Z  * [new branch]                gh/zklaus/22/base       -> origin/gh/zklaus/22/base
2025-12-04T08:53:09.5610717Z  * [new branch]                gh/zklaus/22/head       -> origin/gh/zklaus/22/head
2025-12-04T08:53:09.5610890Z  * [new branch]                gh/zklaus/22/orig       -> origin/gh/zklaus/22/orig
2025-12-04T08:53:09.5611059Z  * [new branch]                gh/zklaus/23/base       -> origin/gh/zklaus/23/base
2025-12-04T08:53:09.5611226Z  * [new branch]                gh/zklaus/23/head       -> origin/gh/zklaus/23/head
2025-12-04T08:53:09.5611393Z  * [new branch]                gh/zklaus/23/orig       -> origin/gh/zklaus/23/orig
2025-12-04T08:53:09.5611561Z  * [new branch]                gh/zklaus/24/base       -> origin/gh/zklaus/24/base
2025-12-04T08:53:09.5611730Z  * [new branch]                gh/zklaus/24/head       -> origin/gh/zklaus/24/head
2025-12-04T08:53:09.5611899Z  * [new branch]                gh/zklaus/24/orig       -> origin/gh/zklaus/24/orig
2025-12-04T08:53:09.5612134Z  * [new branch]                gh/zou3519/1197/base    -> origin/gh/zou3519/1197/base
2025-12-04T08:53:09.5612312Z  * [new branch]                gh/zou3519/1197/head    -> origin/gh/zou3519/1197/head
2025-12-04T08:53:09.5612536Z  * [new branch]                gh/zou3519/1197/orig    -> origin/gh/zou3519/1197/orig
2025-12-04T08:53:09.5612711Z  * [new branch]                gh/zou3519/1199/base    -> origin/gh/zou3519/1199/base
2025-12-04T08:53:09.5612886Z  * [new branch]                gh/zou3519/1199/head    -> origin/gh/zou3519/1199/head
2025-12-04T08:53:09.5613060Z  * [new branch]                gh/zou3519/1199/orig    -> origin/gh/zou3519/1199/orig
2025-12-04T08:53:09.5613234Z  * [new branch]                gh/zou3519/1200/base    -> origin/gh/zou3519/1200/base
2025-12-04T08:53:09.5613408Z  * [new branch]                gh/zou3519/1200/head    -> origin/gh/zou3519/1200/head
2025-12-04T08:53:09.5613581Z  * [new branch]                gh/zou3519/1200/orig    -> origin/gh/zou3519/1200/orig
2025-12-04T08:53:09.5613758Z  * [new branch]                gh/zou3519/1201/base    -> origin/gh/zou3519/1201/base
2025-12-04T08:53:09.5613936Z  * [new branch]                gh/zou3519/1201/head    -> origin/gh/zou3519/1201/head
2025-12-04T08:53:09.5614114Z  * [new branch]                gh/zou3519/1201/orig    -> origin/gh/zou3519/1201/orig
2025-12-04T08:53:09.5614291Z  * [new branch]                gh/zou3519/1202/base    -> origin/gh/zou3519/1202/base
2025-12-04T08:53:09.5614467Z  * [new branch]                gh/zou3519/1202/head    -> origin/gh/zou3519/1202/head
2025-12-04T08:53:09.5614654Z  * [new branch]                gh/zou3519/1202/orig    -> origin/gh/zou3519/1202/orig
2025-12-04T08:53:09.5614830Z  * [new branch]                gh/zpcore/1/base        -> origin/gh/zpcore/1/base
2025-12-04T08:53:09.5615005Z  * [new branch]                gh/zpcore/1/head        -> origin/gh/zpcore/1/head
2025-12-04T08:53:09.5615178Z  * [new branch]                gh/zpcore/11/base       -> origin/gh/zpcore/11/base
2025-12-04T08:53:09.5615352Z  * [new branch]                gh/zpcore/11/head       -> origin/gh/zpcore/11/head
2025-12-04T08:53:09.5615523Z  * [new branch]                gh/zpcore/11/orig       -> origin/gh/zpcore/11/orig
2025-12-04T08:53:09.5615696Z  * [new branch]                gh/zpcore/12/base       -> origin/gh/zpcore/12/base
2025-12-04T08:53:09.5615866Z  * [new branch]                gh/zpcore/12/head       -> origin/gh/zpcore/12/head
2025-12-04T08:53:09.5616037Z  * [new branch]                gh/zpcore/12/orig       -> origin/gh/zpcore/12/orig
2025-12-04T08:53:09.5616206Z  * [new branch]                gh/zpcore/13/base       -> origin/gh/zpcore/13/base
2025-12-04T08:53:09.5620038Z  * [new branch]                gh/zpcore/13/head       -> origin/gh/zpcore/13/head
2025-12-04T08:53:09.5620215Z  * [new branch]                gh/zpcore/13/orig       -> origin/gh/zpcore/13/orig
2025-12-04T08:53:09.5620389Z  * [new branch]                gh/zpcore/14/base       -> origin/gh/zpcore/14/base
2025-12-04T08:53:09.5620599Z  * [new branch]                gh/zpcore/14/head       -> origin/gh/zpcore/14/head
2025-12-04T08:53:09.5620772Z  * [new branch]                gh/zpcore/14/orig       -> origin/gh/zpcore/14/orig
2025-12-04T08:53:09.5620951Z  * [new branch]                gh/zpcore/15/base       -> origin/gh/zpcore/15/base
2025-12-04T08:53:09.5621124Z  * [new branch]                gh/zpcore/15/head       -> origin/gh/zpcore/15/head
2025-12-04T08:53:09.5621297Z  * [new branch]                gh/zpcore/15/orig       -> origin/gh/zpcore/15/orig
2025-12-04T08:53:09.5621471Z  * [new branch]                gh/zpcore/2/base        -> origin/gh/zpcore/2/base
2025-12-04T08:53:09.5621644Z  * [new branch]                gh/zpcore/2/head        -> origin/gh/zpcore/2/head
2025-12-04T08:53:09.5621816Z  * [new branch]                gh/zpcore/21/base       -> origin/gh/zpcore/21/base
2025-12-04T08:53:09.5621989Z  * [new branch]                gh/zpcore/21/head       -> origin/gh/zpcore/21/head
2025-12-04T08:53:09.5622161Z  * [new branch]                gh/zpcore/21/orig       -> origin/gh/zpcore/21/orig
2025-12-04T08:53:09.5622395Z  * [new branch]                gh/zpcore/22/base       -> origin/gh/zpcore/22/base
2025-12-04T08:53:09.5622618Z  * [new branch]                gh/zpcore/22/head       -> origin/gh/zpcore/22/head
2025-12-04T08:53:09.5622790Z  * [new branch]                gh/zpcore/22/orig       -> origin/gh/zpcore/22/orig
2025-12-04T08:53:09.5622961Z  * [new branch]                gh/zpcore/23/base       -> origin/gh/zpcore/23/base
2025-12-04T08:53:09.5623134Z  * [new branch]                gh/zpcore/23/head       -> origin/gh/zpcore/23/head
2025-12-04T08:53:09.5623304Z  * [new branch]                gh/zpcore/23/orig       -> origin/gh/zpcore/23/orig
2025-12-04T08:53:09.5623479Z  * [new branch]                gh/zpcore/24/base       -> origin/gh/zpcore/24/base
2025-12-04T08:53:09.5623652Z  * [new branch]                gh/zpcore/24/head       -> origin/gh/zpcore/24/head
2025-12-04T08:53:09.5623826Z  * [new branch]                gh/zpcore/24/orig       -> origin/gh/zpcore/24/orig
2025-12-04T08:53:09.5624004Z  * [new branch]                gh/zpcore/25/base       -> origin/gh/zpcore/25/base
2025-12-04T08:53:09.5624182Z  * [new branch]                gh/zpcore/25/head       -> origin/gh/zpcore/25/head
2025-12-04T08:53:09.5624354Z  * [new branch]                gh/zpcore/25/orig       -> origin/gh/zpcore/25/orig
2025-12-04T08:53:09.5624526Z  * [new branch]                gh/zpcore/26/base       -> origin/gh/zpcore/26/base
2025-12-04T08:53:09.5624699Z  * [new branch]                gh/zpcore/26/head       -> origin/gh/zpcore/26/head
2025-12-04T08:53:09.5624871Z  * [new branch]                gh/zpcore/26/orig       -> origin/gh/zpcore/26/orig
2025-12-04T08:53:09.5625044Z  * [new branch]                gh/zpcore/27/base       -> origin/gh/zpcore/27/base
2025-12-04T08:53:09.5625216Z  * [new branch]                gh/zpcore/27/head       -> origin/gh/zpcore/27/head
2025-12-04T08:53:09.5625387Z  * [new branch]                gh/zpcore/27/orig       -> origin/gh/zpcore/27/orig
2025-12-04T08:53:09.5625564Z  * [new branch]                gh/zpcore/28/base       -> origin/gh/zpcore/28/base
2025-12-04T08:53:09.5625742Z  * [new branch]                gh/zpcore/28/head       -> origin/gh/zpcore/28/head
2025-12-04T08:53:09.5625914Z  * [new branch]                gh/zpcore/28/orig       -> origin/gh/zpcore/28/orig
2025-12-04T08:53:09.5626086Z  * [new branch]                gh/zpcore/3/base        -> origin/gh/zpcore/3/base
2025-12-04T08:53:09.5626256Z  * [new branch]                gh/zpcore/3/head        -> origin/gh/zpcore/3/head
2025-12-04T08:53:09.5626432Z  * [new branch]                gh/zpcore/4/base        -> origin/gh/zpcore/4/base
2025-12-04T08:53:09.5626604Z  * [new branch]                gh/zpcore/4/head        -> origin/gh/zpcore/4/head
2025-12-04T08:53:09.5626774Z  * [new branch]                gh/zpcore/5/base        -> origin/gh/zpcore/5/base
2025-12-04T08:53:09.5626944Z  * [new branch]                gh/zpcore/5/head        -> origin/gh/zpcore/5/head
2025-12-04T08:53:09.5627115Z  * [new branch]                gh/zpcore/6/base        -> origin/gh/zpcore/6/base
2025-12-04T08:53:09.5627284Z  * [new branch]                gh/zpcore/6/head        -> origin/gh/zpcore/6/head
2025-12-04T08:53:09.5627457Z  * [new branch]                gh/zpcore/7/base        -> origin/gh/zpcore/7/base
2025-12-04T08:53:09.5627625Z  * [new branch]                gh/zpcore/7/head        -> origin/gh/zpcore/7/head
2025-12-04T08:53:09.5627794Z  * [new branch]                gh/zpcore/8/base        -> origin/gh/zpcore/8/base
2025-12-04T08:53:09.5627963Z  * [new branch]                gh/zpcore/8/head        -> origin/gh/zpcore/8/head
2025-12-04T08:53:09.5628134Z  * [new branch]                google-main             -> origin/google-main
2025-12-04T08:53:09.5628323Z  * [new branch]                guangyey/external_stream -> origin/guangyey/external_stream
2025-12-04T08:53:09.5628518Z  * [new branch]                guangyey/test_2025      -> origin/guangyey/test_2025
2025-12-04T08:53:09.5628805Z  * [new branch]                guilhermeleobas/cherry-pick-55d87d9dfd9 -> origin/guilhermeleobas/cherry-pick-55d87d9dfd9
2025-12-04T08:53:09.5629125Z  * [new branch]                hameerabbasi/complex_tensor_subclass -> origin/hameerabbasi/complex_tensor_subclass
2025-12-04T08:53:09.5629412Z  * [new branch]                hameerabbasi/fix-ctensor-gradcheck-tests -> origin/hameerabbasi/fix-ctensor-gradcheck-tests
2025-12-04T08:53:09.5629691Z  * [new branch]                hameerabbasi/gradcheck-allclose -> origin/hameerabbasi/gradcheck-allclose
2025-12-04T08:53:09.5629898Z  * [new branch]                hc_baseline             -> origin/hc_baseline
2025-12-04T08:53:09.5630058Z  * [new branch]                hhh_rand                -> origin/hhh_rand
2025-12-04T08:53:09.5630217Z  * [new branch]                huba/f1                 -> origin/huba/f1
2025-12-04T08:53:09.5630542Z  * [new branch]                increase-timeout-linux-jammy-cuda12_8-py3_10-gcc11-test -> origin/increase-timeout-linux-jammy-cuda12_8-py3_10-gcc11-test
2025-12-04T08:53:09.5630827Z  * [new branch]                inlining                -> origin/inlining
2025-12-04T08:53:09.5630996Z  * [new branch]                inlining-ezyang         -> origin/inlining-ezyang
2025-12-04T08:53:09.5631189Z  * [new branch]                install-torchao-0.13.0  -> origin/install-torchao-0.13.0
2025-12-04T08:53:09.5631520Z  * [new branch]                instrument-trunk-pull-linux-with-job-test-filters -> origin/instrument-trunk-pull-linux-with-job-test-filters
2025-12-04T08:53:09.5631801Z  * [new branch]                invoke-subgraph         -> origin/invoke-subgraph
2025-12-04T08:53:09.5631987Z  * [new branch]                issue#58739             -> origin/issue#58739
2025-12-04T08:53:09.5632166Z  * [new branch]                jainapurva-patch-1      -> origin/jainapurva-patch-1
2025-12-04T08:53:09.5632343Z  * [new branch]                jathu/o3                -> origin/jathu/o3
2025-12-04T08:53:09.5632500Z  * [new branch]                jathu/sve               -> origin/jathu/sve
2025-12-04T08:53:09.5632725Z  * [new branch]                jcaip/test-cusparselt-version-0.6.2 -> origin/jcaip/test-cusparselt-version-0.6.2
2025-12-04T08:53:09.5632991Z  * [new branch]                jcaip/update-cusparselt-0.6.2 -> origin/jcaip/update-cusparselt-0.6.2
2025-12-04T08:53:09.5633242Z  * [new branch]                jiannanWang/memorysnapshot_filter -> origin/jiannanWang/memorysnapshot_filter
2025-12-04T08:53:09.5633495Z  * [new branch]                jiannanWang/profilerstepwarning -> origin/jiannanWang/profilerstepwarning
2025-12-04T08:53:09.5633726Z  * [new branch]                jithunnair-amd-patch-1  -> origin/jithunnair-amd-patch-1
2025-12-04T08:53:09.5633935Z  * [new branch]                jithunnair-amd-patch-10 -> origin/jithunnair-amd-patch-10
2025-12-04T08:53:09.5634139Z  * [new branch]                jithunnair-amd-patch-2  -> origin/jithunnair-amd-patch-2
2025-12-04T08:53:09.5634339Z  * [new branch]                jithunnair-amd-patch-3  -> origin/jithunnair-amd-patch-3
2025-12-04T08:53:09.5634540Z  * [new branch]                jithunnair-amd-patch-4  -> origin/jithunnair-amd-patch-4
2025-12-04T08:53:09.5634742Z  * [new branch]                jithunnair-amd-patch-5  -> origin/jithunnair-amd-patch-5
2025-12-04T08:53:09.5634940Z  * [new branch]                jithunnair-amd-patch-6  -> origin/jithunnair-amd-patch-6
2025-12-04T08:53:09.5635138Z  * [new branch]                jithunnair-amd-patch-7  -> origin/jithunnair-amd-patch-7
2025-12-04T08:53:09.5635333Z  * [new branch]                jithunnair-amd-patch-8  -> origin/jithunnair-amd-patch-8
2025-12-04T08:53:09.5635531Z  * [new branch]                jithunnair-amd-patch-9  -> origin/jithunnair-amd-patch-9
2025-12-04T08:53:09.5635727Z  * [new branch]                justinchu/native-qdq    -> origin/justinchu/native-qdq
2025-12-04T08:53:09.5635918Z  * [new branch]                kainan666/xlf_debug     -> origin/kainan666/xlf_debug
2025-12-04T08:53:09.5636093Z  * [new branch]                kainan_test             -> origin/kainan_test
2025-12-04T08:53:09.5636335Z  * [new branch]                larryliu0820-patch-1    -> origin/larryliu0820-patch-1
2025-12-04T08:53:09.5636611Z  * [new branch]                leslie/test_group_gemm_epilogues -> origin/leslie/test_group_gemm_epilogues
2025-12-04T08:53:09.5636855Z  * [new branch]                lessw2020/fix_cutlass_cache_error -> origin/lessw2020/fix_cutlass_cache_error
2025-12-04T08:53:09.5637073Z  * [new branch]                liaoxuan/shm_all_reduce -> origin/liaoxuan/shm_all_reduce
2025-12-04T08:53:09.5637288Z  * [new branch]                liaoxuan/test_fa_disable_softmax -> origin/liaoxuan/test_fa_disable_softmax
2025-12-04T08:53:09.5637503Z  * [new branch]                liaoxuan/test_int8_sdpa -> origin/liaoxuan/test_int8_sdpa
2025-12-04T08:53:09.5637686Z  * [new branch]                llama4-stable           -> origin/llama4-stable
2025-12-04T08:53:09.5637858Z  * [new branch]                lts/release/1.8         -> origin/lts/release/1.8
2025-12-04T08:53:09.5638041Z  * [new branch]                lucaskabela/#94773      -> origin/lucaskabela/#94773
2025-12-04T08:53:09.5638231Z  * [new branch]                lucaskabela/fix_164876  -> origin/lucaskabela/fix_164876
2025-12-04T08:53:09.5638430Z  * [new branch]                lucaskabela/flop_counter -> origin/lucaskabela/flop_counter
2025-12-04T08:53:09.5638648Z  * [new branch]                lucaskabela/func_under_decomp -> origin/lucaskabela/func_under_decomp
2025-12-04T08:53:09.5638886Z  * [new branch]                lucaskabela/functional_in_dynamo -> origin/lucaskabela/functional_in_dynamo
2025-12-04T08:53:09.5639151Z  * [new branch]                lucaskabela/install_params_as_graph_attr -> origin/lucaskabela/install_params_as_graph_attr
2025-12-04T08:53:09.5639267Z  * [new branch]                lucaskabela/parameters_as_graph_attr -> origin/lucaskabela/parameters_as_graph_attr
2025-12-04T08:53:09.5639399Z  * [new branch]                lucaskabela/remove_aot_dispatcher_metadata -> origin/lucaskabela/remove_aot_dispatcher_metadata
2025-12-04T08:53:09.5639480Z  * [new branch]                lucaskabela/rnn_decomp  -> origin/lucaskabela/rnn_decomp
2025-12-04T08:53:09.5639576Z  * [new branch]                lucaskabela/typing_backends -> origin/lucaskabela/typing_backends
2025-12-04T08:53:09.5639675Z  * [new branch]                lucaskabela/typing_ctx_manager -> origin/lucaskabela/typing_ctx_manager
2025-12-04T08:53:09.5639770Z  * [new branch]                lucaskabela/typing_nn_module -> origin/lucaskabela/typing_nn_module
2025-12-04T08:53:09.5639871Z  * [new branch]                lucaskabela/typing_user_defined -> origin/lucaskabela/typing_user_defined
2025-12-04T08:53:09.5639967Z  * [new branch]                lucaskabela/typing_variables -> origin/lucaskabela/typing_variables
2025-12-04T08:53:09.5640077Z  * [new branch]                lucaskabela/typing_variables_dicts -> origin/lucaskabela/typing_variables_dicts
2025-12-04T08:53:09.5640198Z  * [new branch]                lucaskabela/typing_variables_functions -> origin/lucaskabela/typing_variables_functions
2025-12-04T08:53:09.5640306Z  * [new branch]                lucaskabela/typing_variables_lists -> origin/lucaskabela/typing_variables_lists
2025-12-04T08:53:09.5640381Z  * [new branch]                lw/torch_box_by_ref     -> origin/lw/torch_box_by_ref
2025-12-04T08:53:09.5640471Z  * [new branch]                main                    -> origin/main
2025-12-04T08:53:09.5640541Z  * [new branch]                malfet-patch-1          -> origin/malfet-patch-1
2025-12-04T08:53:09.5640611Z  * [new branch]                malfet-patch-2          -> origin/malfet-patch-2
2025-12-04T08:53:09.5640680Z  * [new branch]                malfet-patch-3          -> origin/malfet-patch-3
2025-12-04T08:53:09.5640745Z  * [new branch]                malfet-patch-4          -> origin/malfet-patch-4
2025-12-04T08:53:09.5640810Z  * [new branch]                malfet-patch-5          -> origin/malfet-patch-5
2025-12-04T08:53:09.5640875Z  * [new branch]                malfet-patch-6          -> origin/malfet-patch-6
2025-12-04T08:53:09.5640981Z  * [new branch]                malfet-patch-7          -> origin/malfet-patch-7
2025-12-04T08:53:09.5641101Z  * [new branch]                malfet-patch-8          -> origin/malfet-patch-8
2025-12-04T08:53:09.5641178Z  * [new branch]                malfet/add-3.14-ci      -> origin/malfet/add-3.14-ci
2025-12-04T08:53:09.5641337Z  * [new branch]                malfet/be-do-not-make-typos-in-build-artifacts -> origin/malfet/be-do-not-make-typos-in-build-artifacts
2025-12-04T08:53:09.5641504Z  * [new branch]                malfet/be-move-more-settings-to-checkout-pytorch -> origin/malfet/be-move-more-settings-to-checkout-pytorch
2025-12-04T08:53:09.5641633Z  * [new branch]                malfet/be-remove-misisng-neon-headers -> origin/malfet/be-remove-misisng-neon-headers
2025-12-04T08:53:09.5641731Z  * [new branch]                malfet/mps-implement-col2im -> origin/malfet/mps-implement-col2im
2025-12-04T08:53:09.5641850Z  * [new branch]                manuel/aoti_metal_shimify-thread_safe -> origin/manuel/aoti_metal_shimify-thread_safe
2025-12-04T08:53:09.5641944Z  * [new branch]                manuel/inductor_link_openmp -> origin/manuel/inductor_link_openmp
2025-12-04T08:53:09.5642018Z  * [new branch]                masnesral/metaconda     -> origin/masnesral/metaconda
2025-12-04T08:53:09.5642094Z  * [new branch]                mem_profiler_flaky_fix  -> origin/mem_profiler_flaky_fix
2025-12-04T08:53:09.5642175Z  * [new branch]                mem_profiler_stack_trace -> origin/mem_profiler_stack_trace
2025-12-04T08:53:09.5642250Z  * [new branch]                memory_profiler_stack   -> origin/memory_profiler_stack
2025-12-04T08:53:09.5642324Z  * [new branch]                metascroy-patch-1       -> origin/metascroy-patch-1
2025-12-04T08:53:09.5642387Z  * [new branch]                mingw_posix             -> origin/mingw_posix
2025-12-04T08:53:09.5642461Z  * [new branch]                mlazos/S429861-debug    -> origin/mlazos/S429861-debug
2025-12-04T08:53:09.5642526Z  * [new branch]                mlazos/aa               -> origin/mlazos/aa
2025-12-04T08:53:09.5642587Z  * [new branch]                mlazos/acts             -> origin/mlazos/acts
2025-12-04T08:53:09.5642660Z  * [new branch]                mlazos/arg-renames      -> origin/mlazos/arg-renames
2025-12-04T08:53:09.5642739Z  * [new branch]                mlazos/bad-cudagraphs   -> origin/mlazos/bad-cudagraphs
2025-12-04T08:53:09.5642840Z  * [new branch]                mlazos/baseline-graph-breaks -> origin/mlazos/baseline-graph-breaks
2025-12-04T08:53:09.5642914Z  * [new branch]                mlazos/beta-tensor      -> origin/mlazos/beta-tensor
2025-12-04T08:53:09.5642980Z  * [new branch]                mlazos/buffers          -> origin/mlazos/buffers
2025-12-04T08:53:09.5643047Z  * [new branch]                mlazos/buffers2         -> origin/mlazos/buffers2
2025-12-04T08:53:09.5643113Z  * [new branch]                mlazos/buffers3         -> origin/mlazos/buffers3
2025-12-04T08:53:09.5643179Z  * [new branch]                mlazos/bwd              -> origin/mlazos/bwd
2025-12-04T08:53:09.5643250Z  * [new branch]                mlazos/combo-test       -> origin/mlazos/combo-test
2025-12-04T08:53:09.5643324Z  * [new branch]                mlazos/ctx-cleanup      -> origin/mlazos/ctx-cleanup
2025-12-04T08:53:09.5643398Z  * [new branch]                mlazos/cuda-cmd-log     -> origin/mlazos/cuda-cmd-log
2025-12-04T08:53:09.5643479Z  * [new branch]                mlazos/cudagraph-tests  -> origin/mlazos/cudagraph-tests
2025-12-04T08:53:09.5643584Z  * [new branch]                mlazos/cudagraphs-measurement -> origin/mlazos/cudagraphs-measurement
2025-12-04T08:53:09.5643657Z  * [new branch]                mlazos/cutlass-test     -> origin/mlazos/cutlass-test
2025-12-04T08:53:09.5643739Z  * [new branch]                mlazos/cutlass-topo-bug -> origin/mlazos/cutlass-topo-bug
2025-12-04T08:53:09.5643820Z  * [new branch]                mlazos/dataclass-proxy  -> origin/mlazos/dataclass-proxy
2025-12-04T08:53:09.5643914Z  * [new branch]                mlazos/dc-attrs         -> origin/mlazos/dc-attrs
2025-12-04T08:53:09.5644007Z  * [new branch]                mlazos/dc-helion        -> origin/mlazos/dc-helion
2025-12-04T08:53:09.5644075Z  * [new branch]                mlazos/dict-fix         -> origin/mlazos/dict-fix
2025-12-04T08:53:09.5644145Z  * [new branch]                mlazos/disable-tf       -> origin/mlazos/disable-tf
2025-12-04T08:53:09.5644211Z  * [new branch]                mlazos/dupe-fix         -> origin/mlazos/dupe-fix
2025-12-04T08:53:09.5644281Z  * [new branch]                mlazos/dyn-batch        -> origin/mlazos/dyn-batch
2025-12-04T08:53:09.5644342Z  * [new branch]                mlazos/evt              -> origin/mlazos/evt
2025-12-04T08:53:09.5644422Z  * [new branch]                mlazos/extract-examples -> origin/mlazos/extract-examples
2025-12-04T08:53:09.5644493Z  * [new branch]                mlazos/foreach-op       -> origin/mlazos/foreach-op
2025-12-04T08:53:09.5644558Z  * [new branch]                mlazos/fp8              -> origin/mlazos/fp8
2025-12-04T08:53:09.5644625Z  * [new branch]                mlazos/fp8-bias         -> origin/mlazos/fp8-bias
2025-12-04T08:53:09.5644707Z  * [new branch]                mlazos/fp8-bias-fusion  -> origin/mlazos/fp8-bias-fusion
2025-12-04T08:53:09.5644776Z  * [new branch]                mlazos/fp8-fixes        -> origin/mlazos/fp8-fixes
2025-12-04T08:53:09.5644841Z  * [new branch]                mlazos/freezing         -> origin/mlazos/freezing
2025-12-04T08:53:09.5644908Z  * [new branch]                mlazos/h-comp           -> origin/mlazos/h-comp
2025-12-04T08:53:09.5644974Z  * [new branch]                mlazos/h-comp2          -> origin/mlazos/h-comp2
2025-12-04T08:53:09.5645040Z  * [new branch]                mlazos/hash-hop         -> origin/mlazos/hash-hop
2025-12-04T08:53:09.5645102Z  * [new branch]                mlazos/hc               -> origin/mlazos/hc
2025-12-04T08:53:09.5645172Z  * [new branch]                mlazos/hc-cycles        -> origin/mlazos/hc-cycles
2025-12-04T08:53:09.5645240Z  * [new branch]                mlazos/hc-fixes         -> origin/mlazos/hc-fixes
2025-12-04T08:53:09.5645309Z  * [new branch]                mlazos/hc-fixes3        -> origin/mlazos/hc-fixes3
2025-12-04T08:53:09.5645376Z  * [new branch]                mlazos/hc-fixes4        -> origin/mlazos/hc-fixes4
2025-12-04T08:53:09.5645441Z  * [new branch]                mlazos/hc-hf            -> origin/mlazos/hc-hf
2025-12-04T08:53:09.5645509Z  * [new branch]                mlazos/hc-mut           -> origin/mlazos/hc-mut
2025-12-04T08:53:09.5645573Z  * [new branch]                mlazos/hc10             -> origin/mlazos/hc10
2025-12-04T08:53:09.5645635Z  * [new branch]                mlazos/hc11             -> origin/mlazos/hc11
2025-12-04T08:53:09.5645694Z  * [new branch]                mlazos/hc12             -> origin/mlazos/hc12
2025-12-04T08:53:09.5645755Z  * [new branch]                mlazos/hc13             -> origin/mlazos/hc13
2025-12-04T08:53:09.5645817Z  * [new branch]                mlazos/hc14             -> origin/mlazos/hc14
2025-12-04T08:53:09.5645877Z  * [new branch]                mlazos/hc15             -> origin/mlazos/hc15
2025-12-04T08:53:09.5645938Z  * [new branch]                mlazos/hc2              -> origin/mlazos/hc2
2025-12-04T08:53:09.5646000Z  * [new branch]                mlazos/hc4              -> origin/mlazos/hc4
2025-12-04T08:53:09.5646059Z  * [new branch]                mlazos/hc5              -> origin/mlazos/hc5
2025-12-04T08:53:09.5646131Z  * [new branch]                mlazos/hc6              -> origin/mlazos/hc6
2025-12-04T08:53:09.5646191Z  * [new branch]                mlazos/hc7              -> origin/mlazos/hc7
2025-12-04T08:53:09.5646249Z  * [new branch]                mlazos/hc8              -> origin/mlazos/hc8
2025-12-04T08:53:09.5646308Z  * [new branch]                mlazos/hc9              -> origin/mlazos/hc9
2025-12-04T08:53:09.5646381Z  * [new branch]                mlazos/hc_baseline2     -> origin/mlazos/hc_baseline2
2025-12-04T08:53:09.5646497Z  * [new branch]                mlazos/inductor-streams -> origin/mlazos/inductor-streams
2025-12-04T08:53:09.5646581Z  * [new branch]                mlazos/main             -> origin/mlazos/main
2025-12-04T08:53:09.5646643Z  * [new branch]                mlazos/mcg2             -> origin/mlazos/mcg2
2025-12-04T08:53:09.5646718Z  * [new branch]                mlazos/meta-guards      -> origin/mlazos/meta-guards
2025-12-04T08:53:09.5646824Z  * [new branch]                mlazos/mlazos/foreach-map-adam -> origin/mlazos/mlazos/foreach-map-adam
2025-12-04T08:53:09.5646923Z  * [new branch]                mlazos/mlazos/tf-mode-backup -> origin/mlazos/mlazos/tf-mode-backup
2025-12-04T08:53:09.5646990Z  * [new branch]                mlazos/mod-fix          -> origin/mlazos/mod-fix
2025-12-04T08:53:09.5647058Z  * [new branch]                mlazos/mode-fix         -> origin/mlazos/mode-fix
2025-12-04T08:53:09.5647124Z  * [new branch]                mlazos/offsets          -> origin/mlazos/offsets
2025-12-04T08:53:09.5647201Z  * [new branch]                mlazos/overguarding     -> origin/mlazos/overguarding
2025-12-04T08:53:09.5647278Z  * [new branch]                mlazos/proxy-ctors      -> origin/mlazos/proxy-ctors
2025-12-04T08:53:09.5647347Z  * [new branch]                mlazos/quant-fix        -> origin/mlazos/quant-fix
2025-12-04T08:53:09.5647417Z  * [new branch]                mlazos/resnet-fix       -> origin/mlazos/resnet-fix
2025-12-04T08:53:09.5647492Z  * [new branch]                mlazos/rm-buf-names     -> origin/mlazos/rm-buf-names
2025-12-04T08:53:09.5647558Z  * [new branch]                mlazos/rm-code          -> origin/mlazos/rm-code
2025-12-04T08:53:09.5647623Z  * [new branch]                mlazos/rm-spam          -> origin/mlazos/rm-spam
2025-12-04T08:53:09.5647685Z  * [new branch]                mlazos/rtp              -> origin/mlazos/rtp
2025-12-04T08:53:09.5647764Z  * [new branch]                mlazos/static-idx-dbg   -> origin/mlazos/static-idx-dbg
2025-12-04T08:53:09.5647851Z  * [new branch]                mlazos/static-inputs-log -> origin/mlazos/static-inputs-log
2025-12-04T08:53:09.5647918Z  * [new branch]                mlazos/stests           -> origin/mlazos/stests
2025-12-04T08:53:09.5647988Z  * [new branch]                mlazos/stream-ops       -> origin/mlazos/stream-ops
2025-12-04T08:53:09.5648054Z  * [new branch]                mlazos/td-fix2          -> origin/mlazos/td-fix2
2025-12-04T08:53:09.5648133Z  * [new branch]                mlazos/tensor-hasattr2  -> origin/mlazos/tensor-hasattr2
2025-12-04T08:53:09.5648195Z  * [new branch]                mlazos/test             -> origin/mlazos/test
2025-12-04T08:53:09.5648259Z  * [new branch]                mlazos/tf-mode          -> origin/mlazos/tf-mode
2025-12-04T08:53:09.5648339Z  * [new branch]                mlazos/tf-mode-backup2  -> origin/mlazos/tf-mode-backup2
2025-12-04T08:53:09.5648417Z  * [new branch]                mlazos/tf-mode-reland   -> origin/mlazos/tf-mode-reland
2025-12-04T08:53:09.5648497Z  * [new branch]                mlazos/tf-mode-reland2  -> origin/mlazos/tf-mode-reland2
2025-12-04T08:53:09.5648576Z  * [new branch]                mlazos/tf-mode-reland3  -> origin/mlazos/tf-mode-reland3
2025-12-04T08:53:09.5648653Z  * [new branch]                mlazos/triton-no-epi    -> origin/mlazos/triton-no-epi
2025-12-04T08:53:09.5648724Z  * [new branch]                mlazos/tune-proto       -> origin/mlazos/tune-proto
2025-12-04T08:53:09.5648798Z  * [new branch]                mlazos/tuple-fixes      -> origin/mlazos/tuple-fixes
2025-12-04T08:53:09.5648873Z  * [new branch]                mlazos/tuple-fixes2     -> origin/mlazos/tuple-fixes2
2025-12-04T08:53:09.5648951Z  * [new branch]                mlazos/tuple-handling   -> origin/mlazos/tuple-handling
2025-12-04T08:53:09.5649031Z  * [new branch]                mlazos/user-stream-base -> origin/mlazos/user-stream-base
2025-12-04T08:53:09.5649104Z  * [new branch]                mlazos/user-streams     -> origin/mlazos/user-streams
2025-12-04T08:53:09.5649227Z  * [new branch]                mlazos/user-streams-backup -> origin/mlazos/user-streams-backup
2025-12-04T08:53:09.5649353Z  * [new branch]                mlazos/user-streams-backup2 -> origin/mlazos/user-streams-backup2
2025-12-04T08:53:09.5649423Z  * [new branch]                mlazos/vary-beta        -> origin/mlazos/vary-beta
2025-12-04T08:53:09.5649493Z  * [new branch]                mlazos/vary-beta2       -> origin/mlazos/vary-beta2
2025-12-04T08:53:09.5649566Z  * [new branch]                mlazos/weird-perf1      -> origin/mlazos/weird-perf1
2025-12-04T08:53:09.5649637Z  * [new branch]                mm_out_dtype_compile    -> origin/mm_out_dtype_compile
2025-12-04T08:53:09.5649702Z  * [new branch]                module-shim             -> origin/module-shim
2025-12-04T08:53:09.5649764Z  * [new branch]                move_config             -> origin/move_config
2025-12-04T08:53:09.5649832Z  * [new branch]                msaroufim/reduce        -> origin/msaroufim/reduce
2025-12-04T08:53:09.5649904Z  * [new branch]                mtia/basic-cmake        -> origin/mtia/basic-cmake
2025-12-04T08:53:09.5650009Z  * [new branch]                mwizak/fix-triton-block-shape -> origin/mwizak/fix-triton-block-shape
2025-12-04T08:53:09.5650076Z  * [new branch]                my_varlen_backup        -> origin/my_varlen_backup
2025-12-04T08:53:09.5650151Z  * [new branch]                nativert_num_outputs    -> origin/nativert_num_outputs
2025-12-04T08:53:09.5650213Z  * [new branch]                new-codegen             -> origin/new-codegen
2025-12-04T08:53:09.5650278Z  * [new branch]                newtest-base            -> origin/newtest-base
2025-12-04T08:53:09.5650350Z  * [new branch]                ngimel/addmm_dtype      -> origin/ngimel/addmm_dtype
2025-12-04T08:53:09.5650443Z  * [new branch]                ngimel/div_inv          -> origin/ngimel/div_inv
2025-12-04T08:53:09.5650520Z  * [new branch]                ngimel/error_index_list -> origin/ngimel/error_index_list
2025-12-04T08:53:09.5650594Z  * [new branch]                ngimel/gather_grid      -> origin/ngimel/gather_grid
2025-12-04T08:53:09.5650684Z  * [new branch]                ngimel/gather_grid_release -> origin/ngimel/gather_grid_release
2025-12-04T08:53:09.5650750Z  * [new branch]                ngimel/gg_new           -> origin/ngimel/gg_new
2025-12-04T08:53:09.5650817Z  * [new branch]                ngimel/hostalloc        -> origin/ngimel/hostalloc
2025-12-04T08:53:09.5650885Z  * [new branch]                ngimel/storage_id       -> origin/ngimel/storage_id
2025-12-04T08:53:09.5650947Z  * [new branch]                nightly                 -> origin/nightly
2025-12-04T08:53:09.5651064Z  * [new branch]                nikitaved/addmm_1_rowcol_lt_path_check -> origin/nikitaved/addmm_1_rowcol_lt_path_check
2025-12-04T08:53:09.5651188Z  * [new branch]                nikitaved/addmm_epilogue_fusions_2d_bias -> origin/nikitaved/addmm_epilogue_fusions_2d_bias
2025-12-04T08:53:09.5651318Z  * [new branch]                nikitaved/addmm_epilogue_fusions_inductor -> origin/nikitaved/addmm_epilogue_fusions_inductor
2025-12-04T08:53:09.5651442Z  * [new branch]                nikitaved/addmm_epilogue_fusions_scratch -> origin/nikitaved/addmm_epilogue_fusions_scratch
2025-12-04T08:53:09.5651558Z  * [new branch]                nikitaved/grad_addmm_epilogue_fusions -> origin/nikitaved/grad_addmm_epilogue_fusions
2025-12-04T08:53:09.5651672Z  * [new branch]                nikitaved/simpler_can_use_32bit_index -> origin/nikitaved/simpler_can_use_32bit_index
2025-12-04T08:53:09.5651741Z  * [new branch]                nikitaved/test          -> origin/nikitaved/test
2025-12-04T08:53:09.5651866Z  * [new branch]                nmacchioni-perf-test-async-autotune -> origin/nmacchioni-perf-test-async-autotune
2025-12-04T08:53:09.5651944Z  * [new branch]                no_distributed_log_spew -> origin/no_distributed_log_spew
2025-12-04T08:53:09.5652008Z  * [new branch]                nofun-hack              -> origin/nofun-hack
2025-12-04T08:53:09.5652109Z  * [new branch]                norm_bench              -> origin/norm_bench
2025-12-04T08:53:09.5652232Z  * [new branch]                nullplay/fuse_matmul    -> origin/nullplay/fuse_matmul
2025-12-04T08:53:09.5652306Z  * [new branch]                nullplay_fuse_matmul    -> origin/nullplay_fuse_matmul
2025-12-04T08:53:09.5652375Z  * [new branch]                optimizer_test          -> origin/optimizer_test
2025-12-04T08:53:09.5652443Z  * [new branch]                orig/release/1.10       -> origin/orig/release/1.10
2025-12-04T08:53:09.5652510Z  * [new branch]                orig/release/1.11       -> origin/orig/release/1.11
2025-12-04T08:53:09.5652579Z  * [new branch]                orig/release/1.12       -> origin/orig/release/1.12
2025-12-04T08:53:09.5652645Z  * [new branch]                orig/release/1.13       -> origin/orig/release/1.13
2025-12-04T08:53:09.5652712Z  * [new branch]                orig/release/1.6        -> origin/orig/release/1.6
2025-12-04T08:53:09.5652780Z  * [new branch]                orig/release/1.7        -> origin/orig/release/1.7
2025-12-04T08:53:09.5652845Z  * [new branch]                orig/release/1.8        -> origin/orig/release/1.8
2025-12-04T08:53:09.5652912Z  * [new branch]                orig/release/1.9        -> origin/orig/release/1.9
2025-12-04T08:53:09.5652979Z  * [new branch]                orig/release/2.0        -> origin/orig/release/2.0
2025-12-04T08:53:09.5653044Z  * [new branch]                orig/release/2.1        -> origin/orig/release/2.1
2025-12-04T08:53:09.5653109Z  * [new branch]                orig/release/2.2        -> origin/orig/release/2.2
2025-12-04T08:53:09.5653175Z  * [new branch]                orig/release/2.3        -> origin/orig/release/2.3
2025-12-04T08:53:09.5653240Z  * [new branch]                orig/release/2.4        -> origin/orig/release/2.4
2025-12-04T08:53:09.5653304Z  * [new branch]                orig/release/2.5        -> origin/orig/release/2.5
2025-12-04T08:53:09.5653372Z  * [new branch]                orig/release/2.6        -> origin/orig/release/2.6
2025-12-04T08:53:09.5653437Z  * [new branch]                orig/release/2.7        -> origin/orig/release/2.7
2025-12-04T08:53:09.5653503Z  * [new branch]                orig/release/2.8        -> origin/orig/release/2.8
2025-12-04T08:53:09.5653568Z  * [new branch]                orig/release/2.9        -> origin/orig/release/2.9
2025-12-04T08:53:09.5653654Z  * [new branch]                origin/gh/fxdawnn/1/base -> origin/origin/gh/fxdawnn/1/base
2025-12-04T08:53:09.5653738Z  * [new branch]                origin/gh/fxdawnn/1/orig -> origin/origin/gh/fxdawnn/1/orig
2025-12-04T08:53:09.5653820Z  * [new branch]                origin/gh/zpcore/14/orig -> origin/origin/gh/zpcore/14/orig
2025-12-04T08:53:09.5653889Z  * [new branch]                oulgen-patch-1          -> origin/oulgen-patch-1
2025-12-04T08:53:09.5653958Z  * [new branch]                oulgen-patch-2          -> origin/oulgen-patch-2
2025-12-04T08:53:09.5654026Z  * [new branch]                oulgen-patch-3          -> origin/oulgen-patch-3
2025-12-04T08:53:09.5654093Z  * [new branch]                oulgen-patch-4          -> origin/oulgen-patch-4
2025-12-04T08:53:09.5654163Z  * [new branch]                padded-tensor           -> origin/padded-tensor
2025-12-04T08:53:09.5654228Z  * [new branch]                pca2                    -> origin/pca2
2025-12-04T08:53:09.5654299Z  * [new branch]                per_channel_backup      -> origin/per_channel_backup
2025-12-04T08:53:09.5654365Z  * [new branch]                perf_ops                -> origin/perf_ops
2025-12-04T08:53:09.5654429Z  * [new branch]                perf_ops_2_9            -> origin/perf_ops_2_9
2025-12-04T08:53:09.5654500Z  * [new branch]                pianpwk-patch-1         -> origin/pianpwk-patch-1
2025-12-04T08:53:09.5654588Z  * [new branch]                pianpwk/__draft_debug_mode -> origin/pianpwk/__draft_debug_mode
2025-12-04T08:53:09.5654699Z  * [new branch]                pianpwk/_debug_mode_for_triton_draft -> origin/pianpwk/_debug_mode_for_triton_draft
2025-12-04T08:53:09.5654837Z  * [new branch]                pianpwk/_debug_nn_module_compile -> origin/pianpwk/_debug_nn_module_compile
2025-12-04T08:53:09.5654947Z  * [new branch]                pianpwk/_draft_triton_11_3 -> origin/pianpwk/_draft_triton_11_3
2025-12-04T08:53:09.5655039Z  * [new branch]                pianpwk/_manual_bucket_draft -> origin/pianpwk/_manual_bucket_draft
2025-12-04T08:53:09.5655142Z  * [new branch]                pianpwk/_profile_w_dispatch_keys -> origin/pianpwk/_profile_w_dispatch_keys
2025-12-04T08:53:09.5655240Z  * [new branch]                pianpwk/_super_draft_debug_mode -> origin/pianpwk/_super_draft_debug_mode
2025-12-04T08:53:09.5655345Z  * [new branch]                pianpwk/_unbacked_local_shard_size -> origin/pianpwk/_unbacked_local_shard_size
2025-12-04T08:53:09.5655421Z  * [new branch]                pianpwk/anomaly_tb      -> origin/pianpwk/anomaly_tb
2025-12-04T08:53:09.5655504Z  * [new branch]                pianpwk/auto_fx_annotate -> origin/pianpwk/auto_fx_annotate
2025-12-04T08:53:09.5655617Z  * [new branch]                pianpwk/backed_size_oblivious_export -> origin/pianpwk/backed_size_oblivious_export
2025-12-04T08:53:09.5655705Z  * [new branch]                pianpwk/bert_dynamic_perf -> origin/pianpwk/bert_dynamic_perf
2025-12-04T08:53:09.5655801Z  * [new branch]                pianpwk/debug_fwd_stack_traces -> origin/pianpwk/debug_fwd_stack_traces
2025-12-04T08:53:09.5655886Z  * [new branch]                pianpwk/debug_hash_tensor -> origin/pianpwk/debug_hash_tensor
2025-12-04T08:53:09.5655977Z  * [new branch]                pianpwk/debug_mode_annotate -> origin/pianpwk/debug_mode_annotate
2025-12-04T08:53:09.5656065Z  * [new branch]                pianpwk/debug_mode_defaults -> origin/pianpwk/debug_mode_defaults
2025-12-04T08:53:09.5656146Z  * [new branch]                pianpwk/debug_mode_hacks -> origin/pianpwk/debug_mode_hacks
2025-12-04T08:53:09.5656256Z  * [new branch]                pianpwk/debug_mode_opcall_refactor -> origin/pianpwk/debug_mode_opcall_refactor
2025-12-04T08:53:09.5656344Z  * [new branch]                pianpwk/debug_mode_show_ids -> origin/pianpwk/debug_mode_show_ids
2025-12-04T08:53:09.5656430Z  * [new branch]                pianpwk/debug_mode_triton -> origin/pianpwk/debug_mode_triton
2025-12-04T08:53:09.5656526Z  * [new branch]                pianpwk/debug_show_stack_trace -> origin/pianpwk/debug_show_stack_trace
2025-12-04T08:53:09.5656626Z  * [new branch]                pianpwk/debug_wait_on_collective -> origin/pianpwk/debug_wait_on_collective
2025-12-04T08:53:09.5656723Z  * [new branch]                pianpwk/debugmode_compile_tf -> origin/pianpwk/debugmode_compile_tf
2025-12-04T08:53:09.5656849Z  * [new branch]                pianpwk/dispatch_key_debugging_for_debug -> origin/pianpwk/dispatch_key_debugging_for_debug
2025-12-04T08:53:09.5656955Z  * [new branch]                pianpwk/draft_debug_mode_tfcompile -> origin/pianpwk/draft_debug_mode_tfcompile
2025-12-04T08:53:09.5657051Z  * [new branch]                pianpwk/draft_multikernel_nn -> origin/pianpwk/draft_multikernel_nn
2025-12-04T08:53:09.5657167Z  * [new branch]                pianpwk/draft_multikernel_status_10_5 -> origin/pianpwk/draft_multikernel_status_10_5
2025-12-04T08:53:09.5657259Z  * [new branch]                pianpwk/dtensor_custom_chunk -> origin/pianpwk/dtensor_custom_chunk
2025-12-04T08:53:09.5657364Z  * [new branch]                pianpwk/dtensor_unbacked_keypath -> origin/pianpwk/dtensor_unbacked_keypath
2025-12-04T08:53:09.5657443Z  * [new branch]                pianpwk/event_list_tree -> origin/pianpwk/event_list_tree
2025-12-04T08:53:09.5657523Z  * [new branch]                pianpwk/false_numel_refs -> origin/pianpwk/false_numel_refs
2025-12-04T08:53:09.5657602Z  * [new branch]                pianpwk/maybe_guard_rel -> origin/pianpwk/maybe_guard_rel
2025-12-04T08:53:09.5657707Z  * [new branch]                pianpwk/multikernel_hints_draft -> origin/pianpwk/multikernel_hints_draft
2025-12-04T08:53:09.5657848Z  * [new branch]                pianpwk/no_size_oblivious_slice_scat -> origin/pianpwk/no_size_oblivious_slice_scat
2025-12-04T08:53:09.5657991Z  * [new branch]                pianpwk/oblivious_reshape_view_better -> origin/pianpwk/oblivious_reshape_view_better
2025-12-04T08:53:09.5658074Z  * [new branch]                pianpwk/pre_forward_hook -> origin/pianpwk/pre_forward_hook
2025-12-04T08:53:09.5658182Z  * [new branch]                pianpwk/skip_python_keys_alternate -> origin/pianpwk/skip_python_keys_alternate
2025-12-04T08:53:09.5658286Z  * [new branch]                pianpwk/skip_python_keys_in_guards -> origin/pianpwk/skip_python_keys_in_guards
2025-12-04T08:53:09.5658368Z  * [new branch]                pianpwk/sym_tokens_draft -> origin/pianpwk/sym_tokens_draft
2025-12-04T08:53:09.5658449Z  * [new branch]                pianpwk/symint_one_hot  -> origin/pianpwk/symint_one_hot
2025-12-04T08:53:09.5658561Z  * [new branch]                pianpwk/test_pointwise_guard_or_false -> origin/pianpwk/test_pointwise_guard_or_false
2025-12-04T08:53:09.5658662Z  * [new branch]                pianpwk/totally_draft_sym_wrap -> origin/pianpwk/totally_draft_sym_wrap
2025-12-04T08:53:09.5658746Z  * [new branch]                pianpwk/try_dumb_stuff  -> origin/pianpwk/try_dumb_stuff
2025-12-04T08:53:09.5658825Z  * [new branch]                pianpwk/try_dumb_stuff_2 -> origin/pianpwk/try_dumb_stuff_2
2025-12-04T08:53:09.5658916Z  * [new branch]                pianpwk/unbacked_dtensor_mm -> origin/pianpwk/unbacked_dtensor_mm
2025-12-04T08:53:09.5659013Z  * [new branch]                pianpwk/unbacked_tracing_12_2 -> origin/pianpwk/unbacked_tracing_12_2
2025-12-04T08:53:09.5659088Z  * [new branch]                pianpwk/user_symints    -> origin/pianpwk/user_symints
2025-12-04T08:53:09.5659167Z  * [new branch]                pianpwk/wan21_reshape   -> origin/pianpwk/wan21_reshape
2025-12-04T08:53:09.5659261Z  * [new branch]                piz/fix_partial_backward_1112 -> origin/piz/fix_partial_backward_1112
2025-12-04T08:53:09.5659337Z  * [new branch]                piz/prop_cache_clean    -> origin/piz/prop_cache_clean
2025-12-04T08:53:09.5659407Z  * [new branch]                pool-separate           -> origin/pool-separate
2025-12-04T08:53:09.5659469Z  * [new branch]                pr-156087               -> origin/pr-156087
2025-12-04T08:53:09.5659529Z  * [new branch]                pr/131860               -> origin/pr/131860
2025-12-04T08:53:09.5659599Z  * [new branch]                predispatch_to          -> origin/predispatch_to
2025-12-04T08:53:09.5659662Z  * [new branch]                protect-c17             -> origin/protect-c17
2025-12-04T08:53:09.5659729Z  * [new branch]                pt-opt-cuda3            -> origin/pt-opt-cuda3
2025-12-04T08:53:09.5659810Z  * [new branch]                python_compiled_autograd -> origin/python_compiled_autograd
2025-12-04T08:53:09.5659938Z  * [new branch]                q1l1/fix_device_moved_constant_type_unknown -> origin/q1l1/fix_device_moved_constant_type_unknown
2025-12-04T08:53:09.5660077Z  * [new branch]                q1l1/fix_wrong_default_type_for_kernel_call_args -> origin/q1l1/fix_wrong_default_type_for_kernel_call_args
2025-12-04T08:53:09.5660160Z  * [new branch]                qchip/export-D54134695  -> origin/qchip/export-D54134695
2025-12-04T08:53:09.5660234Z  * [new branch]                quote-pytest_cache      -> origin/quote-pytest_cache
2025-12-04T08:53:09.5660331Z  * [new branch]                reland-accgrad-stream-warn -> origin/reland-accgrad-stream-warn
2025-12-04T08:53:09.5660396Z  * [new branch]                release/1.10            -> origin/release/1.10
2025-12-04T08:53:09.5660490Z  * [new branch]                release/1.11            -> origin/release/1.11
2025-12-04T08:53:09.5660553Z  * [new branch]                release/1.12            -> origin/release/1.12
2025-12-04T08:53:09.5660615Z  * [new branch]                release/1.13            -> origin/release/1.13
2025-12-04T08:53:09.5660742Z  * [new branch]                release/1.4             -> origin/release/1.4
2025-12-04T08:53:09.5660805Z  * [new branch]                release/1.4.1           -> origin/release/1.4.1
2025-12-04T08:53:09.5660911Z  * [new branch]                release/1.5             -> origin/release/1.5
2025-12-04T08:53:09.5660972Z  * [new branch]                release/1.6             -> origin/release/1.6
2025-12-04T08:53:09.5661033Z  * [new branch]                release/1.7             -> origin/release/1.7
2025-12-04T08:53:09.5661093Z  * [new branch]                release/1.8             -> origin/release/1.8
2025-12-04T08:53:09.5661153Z  * [new branch]                release/1.9             -> origin/release/1.9
2025-12-04T08:53:09.5661213Z  * [new branch]                release/2.0             -> origin/release/2.0
2025-12-04T08:53:09.5661273Z  * [new branch]                release/2.1             -> origin/release/2.1
2025-12-04T08:53:09.5661332Z  * [new branch]                release/2.2             -> origin/release/2.2
2025-12-04T08:53:09.5661395Z  * [new branch]                release/2.3             -> origin/release/2.3
2025-12-04T08:53:09.5661455Z  * [new branch]                release/2.4             -> origin/release/2.4
2025-12-04T08:53:09.5661516Z  * [new branch]                release/2.5             -> origin/release/2.5
2025-12-04T08:53:09.5661576Z  * [new branch]                release/2.6             -> origin/release/2.6
2025-12-04T08:53:09.5661635Z  * [new branch]                release/2.7             -> origin/release/2.7
2025-12-04T08:53:09.5661695Z  * [new branch]                release/2.8             -> origin/release/2.8
2025-12-04T08:53:09.5661755Z  * [new branch]                release/2.9             -> origin/release/2.9
2025-12-04T08:53:09.5661819Z  * [new branch]                release_notes           -> origin/release_notes
2025-12-04T08:53:09.5661894Z  * [new branch]                remove_pyinterpreter    -> origin/remove_pyinterpreter
2025-12-04T08:53:09.5662021Z  * [new branch]                replace-pytorch-labs-20250812-195836 -> origin/replace-pytorch-labs-20250812-195836
2025-12-04T08:53:09.5662145Z  * [new branch]                replace-pytorch-labs-20250812-200248 -> origin/replace-pytorch-labs-20250812-200248
2025-12-04T08:53:09.5662267Z  * [new branch]                replace-pytorch-labs-20250812-200324 -> origin/replace-pytorch-labs-20250812-200324
2025-12-04T08:53:09.5662387Z  * [new branch]                replace-pytorch-labs-20250812-204020 -> origin/replace-pytorch-labs-20250812-204020
2025-12-04T08:53:09.5662517Z  * [new branch]                revert-131069-gh/krzysztofjordan/1/head -> origin/revert-131069-gh/krzysztofjordan/1/head
2025-12-04T08:53:09.5662629Z  * [new branch]                revert-131469-gh/andrewor14/51/head -> origin/revert-131469-gh/andrewor14/51/head
2025-12-04T08:53:09.5662733Z  * [new branch]                revert-152361-gh/fadara01/1/head -> origin/revert-152361-gh/fadara01/1/head
2025-12-04T08:53:09.5662835Z  * [new branch]                revert-156870-gh/skarjala/3/head -> origin/revert-156870-gh/skarjala/3/head
2025-12-04T08:53:09.5663012Z  * [new branch]                revert-157914-cherry-pick-157503-by-pytorch_bot_bot_ -> origin/revert-157914-cherry-pick-157503-by-pytorch_bot_bot_
2025-12-04T08:53:09.5663110Z  * [new branch]                revert-hoo-invoke-subgraph -> origin/revert-hoo-invoke-subgraph
2025-12-04T08:53:09.5663208Z  * [new branch]                revert_always_build_distributed -> origin/revert_always_build_distributed
2025-12-04T08:53:09.5663276Z  * [new branch]                rms_norm_patch          -> origin/rms_norm_patch
2025-12-04T08:53:09.5663372Z  * [new branch]                ruisi/fix_all_to_all_estimation -> origin/ruisi/fix_all_to_all_estimation
2025-12-04T08:53:09.5663456Z  * [new branch]                ruisi/fix_comm_estimation -> origin/ruisi/fix_comm_estimation
2025-12-04T08:53:09.5663563Z  * [new branch]                ruisi/fix_dynamic_shape_estimation -> origin/ruisi/fix_dynamic_shape_estimation
2025-12-04T08:53:09.5663683Z  * [new branch]                ruisi/fix_llama3_autobucketing -> origin/ruisi/fix_llama3_autobucketing
2025-12-04T08:53:09.5663816Z  * [new branch]                ruisi/fix_manual_bucketing_ep_pass -> origin/ruisi/fix_manual_bucketing_ep_pass
2025-12-04T08:53:09.5663900Z  * [new branch]                ruisi/manual_bucket_pass -> origin/ruisi/manual_bucket_pass
2025-12-04T08:53:09.5664047Z  * [new branch]                ryanguo99/cleanup-dynamo-expected-failures -> origin/ryanguo99/cleanup-dynamo-expected-failures
2025-12-04T08:53:09.5664134Z  * [new branch]                ryanguo99/fix-closure-var -> origin/ryanguo99/fix-closure-var
2025-12-04T08:53:09.5664211Z  * [new branch]                rzou/faketensor_bench   -> origin/rzou/faketensor_bench
2025-12-04T08:53:09.5664273Z  * [new branch]                rzou/njt                -> origin/rzou/njt
2025-12-04T08:53:09.5664336Z  * [new branch]                rzou/pca                -> origin/rzou/pca
2025-12-04T08:53:09.5664401Z  * [new branch]                rzou/realprop           -> origin/rzou/realprop
2025-12-04T08:53:09.5664467Z  * [new branch]                samplevllm              -> origin/samplevllm
2025-12-04T08:53:09.5664636Z  * [new branch]                sanchitintel/weird_thing_with_test_cpu_select_algorithm -> origin/sanchitintel/weird_thing_with_test_cpu_select_algorithm
2025-12-04T08:53:09.5664729Z  * [new branch]                sapling-pr-archive-SS-JIA -> origin/sapling-pr-archive-SS-JIA
2025-12-04T08:53:09.5664842Z  * [new branch]                sapling-pr-archive-tushar00jain -> origin/sapling-pr-archive-tushar00jain
2025-12-04T08:53:09.5664902Z  * [new branch]                save                    -> origin/save
2025-12-04T08:53:09.5664962Z  * [new branch]                scaled_mm               -> origin/scaled_mm
2025-12-04T08:53:09.5665025Z  * [new branch]                scan_attempt            -> origin/scan_attempt
2025-12-04T08:53:09.5665087Z  * [new branch]                sdym/2.5.1              -> origin/sdym/2.5.1
2025-12-04T08:53:09.5665195Z  * [new branch]                sekyondaMeta-dynamoconfig-fix -> origin/sekyondaMeta-dynamoconfig-fix
2025-12-04T08:53:09.5665271Z  * [new branch]                shengf/fx-xform-perf    -> origin/shengf/fx-xform-perf
2025-12-04T08:53:09.5665351Z  * [new branch]                shoumikhin-patch-1      -> origin/shoumikhin-patch-1
2025-12-04T08:53:09.5665426Z  * [new branch]                solve-accuracy-fix      -> origin/solve-accuracy-fix
2025-12-04T08:53:09.5665505Z  * [new branch]                some_rocm_inductor_skips -> origin/some_rocm_inductor_skips
2025-12-04T08:53:09.5665587Z  * [new branch]                soulitzer/stash-tls-ac  -> origin/soulitzer/stash-tls-ac
2025-12-04T08:53:09.5665669Z  * [new branch]                sparse-mm-bf16-support  -> origin/sparse-mm-bf16-support
2025-12-04T08:53:09.5665743Z  * [new branch]                starterTaskUpdate       -> origin/starterTaskUpdate
2025-12-04T08:53:09.5665802Z  * [new branch]                suo                     -> origin/suo
2025-12-04T08:53:09.5665864Z  * [new branch]                sve-poc                 -> origin/sve-poc
2025-12-04T08:53:09.5665926Z  * [new branch]                switch-bn               -> origin/switch-bn
2025-12-04T08:53:09.5666019Z  * [new branch]                sy_annotation_in_autograd_hop -> origin/sy_annotation_in_autograd_hop
2025-12-04T08:53:09.5666087Z  * [new branch]                sy_aot_eager_record     -> origin/sy_aot_eager_record
2025-12-04T08:53:09.5666157Z  * [new branch]                sy_custom_bucketing     -> origin/sy_custom_bucketing
2025-12-04T08:53:09.5666224Z  * [new branch]                sy_debug_mode_test      -> origin/sy_debug_mode_test
2025-12-04T08:53:09.5666288Z  * [new branch]                sy_deserialize          -> origin/sy_deserialize
2025-12-04T08:53:09.5666354Z  * [new branch]                sy_dump_gm_code         -> origin/sy_dump_gm_code
2025-12-04T08:53:09.5666414Z  * [new branch]                sy_exp                  -> origin/sy_exp
2025-12-04T08:53:09.5666514Z  * [new branch]                sy_export_annotation    -> origin/sy_export_annotation
2025-12-04T08:53:09.5666583Z  * [new branch]                sy_invoke_subgraph      -> origin/sy_invoke_subgraph
2025-12-04T08:53:09.5666675Z  * [new branch]                sy_kernel_bw_name       -> origin/sy_kernel_bw_name
2025-12-04T08:53:09.5666736Z  * [new branch]                sy_multi_arch           -> origin/sy_multi_arch
2025-12-04T08:53:09.5666804Z  * [new branch]                sy_nn_module_stack      -> origin/sy_nn_module_stack
2025-12-04T08:53:09.5666874Z  * [new branch]                sy_original_dtensor     -> origin/sy_original_dtensor
2025-12-04T08:53:09.5666940Z  * [new branch]                sy_profiler_cia         -> origin/sy_profiler_cia
2025-12-04T08:53:09.5667003Z  * [new branch]                symm_mem_sync           -> origin/symm_mem_sync
2025-12-04T08:53:09.5667086Z  * [new branch]                sympy-bottleneck-repro  -> origin/sympy-bottleneck-repro
2025-12-04T08:53:09.5667163Z  * [new branch]                tensordict_integration  -> origin/tensordict_integration
2025-12-04T08:53:09.5667246Z  * [new branch]                test-move-conda-builds  -> origin/test-move-conda-builds
2025-12-04T08:53:09.5667309Z  * [new branch]                test-old                -> origin/test-old
2025-12-04T08:53:09.5667372Z  * [new branch]                test/bmm_heur           -> origin/test/bmm_heur
2025-12-04T08:53:09.5667470Z  * [new branch]                tianren/customOp_autotune_fix -> origin/tianren/customOp_autotune_fix
2025-12-04T08:53:09.5667582Z  * [new branch]                tianren/customOp_enable_max_autotune -> origin/tianren/customOp_enable_max_autotune
2025-12-04T08:53:09.5667664Z  * [new branch]                tianren/customOp_fusion -> origin/tianren/customOp_fusion
2025-12-04T08:53:09.5667788Z  * [new branch]                tianren/customop_collectiveop_benchmark -> origin/tianren/customop_collectiveop_benchmark
2025-12-04T08:53:09.5667923Z  * [new branch]                tianren/customop_collectiveop_benchmark_fix -> origin/tianren/customop_collectiveop_benchmark_fix
2025-12-04T08:53:09.5668027Z  * [new branch]                tianren/customop_dynamic_config -> origin/tianren/customop_dynamic_config
2025-12-04T08:53:09.5668120Z  * [new branch]                tianren/dynamic_range_input -> origin/tianren/dynamic_range_input
2025-12-04T08:53:09.5668220Z  * [new branch]                tianren/dynamic_range_input_fix -> origin/tianren/dynamic_range_input_fix
2025-12-04T08:53:09.5668324Z  * [new branch]                tianren/dynamic_range_input_merge -> origin/tianren/dynamic_range_input_merge
2025-12-04T08:53:09.5668424Z  * [new branch]                tianren/flex_paged_attn_fix_temp -> origin/tianren/flex_paged_attn_fix_temp
2025-12-04T08:53:09.5668502Z  * [new branch]                tianren/fx_codegen_dump -> origin/tianren/fx_codegen_dump
2025-12-04T08:53:09.5668586Z  * [new branch]                tianren/symmetric_memory -> origin/tianren/symmetric_memory
2025-12-04T08:53:09.5668651Z  * [new branch]                tianren/test            -> origin/tianren/test
2025-12-04T08:53:09.5668727Z  * [new branch]                tidy_performance_cyy    -> origin/tidy_performance_cyy
2025-12-04T08:53:09.5668788Z  * [new branch]                tmp                     -> origin/tmp
2025-12-04T08:53:09.5668853Z  * [new branch]                torchtitan_ep           -> origin/torchtitan_ep
2025-12-04T08:53:09.5668930Z  * [new branch]                torchtitan_integration  -> origin/torchtitan_integration
2025-12-04T08:53:09.5669013Z  * [new branch]                trace_fsdp_torchtune_lora -> origin/trace_fsdp_torchtune_lora
2025-12-04T08:53:09.5669096Z  * [new branch]                traceable_fsdp_unit_tests -> origin/traceable_fsdp_unit_tests
2025-12-04T08:53:09.5669166Z  * [new branch]                tree_loop_vec_base      -> origin/tree_loop_vec_base
2025-12-04T08:53:09.5669230Z  * [new branch]                triton_kernel           -> origin/triton_kernel
2025-12-04T08:53:09.5669291Z  * [new branch]                tt_pkg_1908             -> origin/tt_pkg_1908
2025-12-04T08:53:09.5669385Z  * [new branch]                type_dec                -> origin/type_dec
2025-12-04T08:53:09.5669498Z  * [new branch]                udate-sphinx-dependancies -> origin/udate-sphinx-dependancies
2025-12-04T08:53:09.5669637Z  * [new branch]                update-audio-commit-hash/17630256502-1803-1 -> origin/update-audio-commit-hash/17630256502-1803-1
2025-12-04T08:53:09.5669773Z  * [new branch]                update-audio-commit-hash/19087141161-1916-1 -> origin/update-audio-commit-hash/19087141161-1916-1
2025-12-04T08:53:09.5669907Z  * [new branch]                update-audio-commit-hash/19250643381-1929-1 -> origin/update-audio-commit-hash/19250643381-1929-1
2025-12-04T08:53:09.5670041Z  * [new branch]                update-audio-commit-hash/19397724337-1935-1 -> origin/update-audio-commit-hash/19397724337-1935-1
2025-12-04T08:53:09.5670176Z  * [new branch]                update-audio-commit-hash/19555670148-1941-1 -> origin/update-audio-commit-hash/19555670148-1941-1
2025-12-04T08:53:09.5670311Z  * [new branch]                update-audio-commit-hash/19750627930-1946-1 -> origin/update-audio-commit-hash/19750627930-1946-1
2025-12-04T08:53:09.5670474Z  * [new branch]                update-triton-commit-hash/13663274526-1487-2 -> origin/update-triton-commit-hash/13663274526-1487-2
2025-12-04T08:53:09.5670612Z  * [new branch]                update-vision-commit-hash/19087141161-1916-1 -> origin/update-vision-commit-hash/19087141161-1916-1
2025-12-04T08:53:09.5670747Z  * [new branch]                update-vision-commit-hash/19184897099-1925-1 -> origin/update-vision-commit-hash/19184897099-1925-1
2025-12-04T08:53:09.5670884Z  * [new branch]                update-vision-commit-hash/19250643381-1929-1 -> origin/update-vision-commit-hash/19250643381-1929-1
2025-12-04T08:53:09.5671018Z  * [new branch]                update-vision-commit-hash/19381328640-1934-1 -> origin/update-vision-commit-hash/19381328640-1934-1
2025-12-04T08:53:09.5671153Z  * [new branch]                update-vision-commit-hash/19485237164-1938-1 -> origin/update-vision-commit-hash/19485237164-1938-1
2025-12-04T08:53:09.5671286Z  * [new branch]                update-vllm-commit-hash/18451675449-1879-1 -> origin/update-vllm-commit-hash/18451675449-1879-1
2025-12-04T08:53:09.5671369Z  * [new branch]                update-vllm-dockerfile  -> origin/update-vllm-dockerfile
2025-12-04T08:53:09.5671494Z  * [new branch]                update-xla-commit-hash/19224287370-211-1 -> origin/update-xla-commit-hash/19224287370-211-1
2025-12-04T08:53:09.5671620Z  * [new branch]                update-xla-commit-hash/19422028566-212-1 -> origin/update-xla-commit-hash/19422028566-212-1
2025-12-04T08:53:09.5671742Z  * [new branch]                update-xla-commit-hash/19626841311-213-1 -> origin/update-xla-commit-hash/19626841311-213-1
2025-12-04T08:53:09.5671871Z  * [new branch]                update_docs_torch_multinomial_issue#125388 -> origin/update_docs_torch_multinomial_issue#125388
2025-12-04T08:53:09.5671952Z  * [new branch]                update_operator_readme  -> origin/update_operator_readme
2025-12-04T08:53:09.5672042Z  * [new branch]                update_slow_tests_1722488736 -> origin/update_slow_tests_1722488736
2025-12-04T08:53:09.5672130Z  * [new branch]                update_slow_tests_1722879173 -> origin/update_slow_tests_1722879173
2025-12-04T08:53:09.5672216Z  * [new branch]                update_slow_tests_1762155677 -> origin/update_slow_tests_1762155677
2025-12-04T08:53:09.5672302Z  * [new branch]                update_slow_tests_1763365283 -> origin/update_slow_tests_1763365283
2025-12-04T08:53:09.5672388Z  * [new branch]                update_submodule_FBGEMM -> origin/update_submodule_FBGEMM
2025-12-04T08:53:09.5672465Z  * [new branch]                update_submodule_kineto -> origin/update_submodule_kineto
2025-12-04T08:53:09.5672555Z  * [new branch]                update_submodule_tensorpipe -> origin/update_submodule_tensorpipe
2025-12-04T08:53:09.5672706Z  * [new branch]                upload-tests-for-autorevert -> origin/upload-tests-for-autorevert
2025-12-04T08:53:09.5672769Z  * [new branch]                v0.1.2                  -> origin/v0.1.2
2025-12-04T08:53:09.5672867Z  * [new branch]                v1.0.1                  -> origin/v1.0.1
2025-12-04T08:53:09.5672926Z  * [new branch]                v1.0.3                  -> origin/v1.0.3
2025-12-04T08:53:09.5672983Z  * [new branch]                v1.1.0                  -> origin/v1.1.0
2025-12-04T08:53:09.5673039Z  * [new branch]                v1.2.0                  -> origin/v1.2.0
2025-12-04T08:53:09.5673096Z  * [new branch]                v1.3.0                  -> origin/v1.3.0
2025-12-04T08:53:09.5673151Z  * [new branch]                v1.3.1                  -> origin/v1.3.1
2025-12-04T08:53:09.5673214Z  * [new branch]                validate_fn             -> origin/validate_fn
2025-12-04T08:53:09.5673283Z  * [new branch]                validations_2.6         -> origin/validations_2.6
2025-12-04T08:53:09.5673351Z  * [new branch]                validations_2.8         -> origin/validations_2.8
2025-12-04T08:53:09.5673415Z  * [new branch]                varlen-api              -> origin/varlen-api
2025-12-04T08:53:09.5673492Z  * [new branch]                varlen-api-backup       -> origin/varlen-api-backup
2025-12-04T08:53:09.5673568Z  * [new branch]                varlen_batch_invariance -> origin/varlen_batch_invariance
2025-12-04T08:53:09.5673634Z  * [new branch]                viable/strict           -> origin/viable/strict
2025-12-04T08:53:09.5673751Z  * [new branch]                vishal9-team/dtensor_parallelism_toy -> origin/vishal9-team/dtensor_parallelism_toy
2025-12-04T08:53:09.5673815Z  * [new branch]                vllmbuildci             -> origin/vllmbuildci
2025-12-04T08:53:09.5673876Z  * [new branch]                vllmpin                 -> origin/vllmpin
2025-12-04T08:53:09.5673964Z  * [new branch]                vscode-recommend-pyrefly -> origin/vscode-recommend-pyrefly
2025-12-04T08:53:09.5674031Z  * [new branch]                wdvr-patch-1            -> origin/wdvr-patch-1
2025-12-04T08:53:09.5674097Z  * [new branch]                wdvr/iss_145259         -> origin/wdvr/iss_145259
2025-12-04T08:53:09.5674159Z  * [new branch]                whc/pei                 -> origin/whc/pei
2025-12-04T08:53:09.5674223Z  * [new branch]                whc/pp_fix              -> origin/whc/pp_fix
2025-12-04T08:53:09.5674287Z  * [new branch]                whc/sharding            -> origin/whc/sharding
2025-12-04T08:53:09.5674350Z  * [new branch]                whc/sharding2           -> origin/whc/sharding2
2025-12-04T08:53:09.5674410Z  * [new branch]                whc/uneven              -> origin/whc/uneven
2025-12-04T08:53:09.5674482Z  * [new branch]                whc/uneven-merge        -> origin/whc/uneven-merge
2025-12-04T08:53:09.5674543Z  * [new branch]                win_warnings            -> origin/win_warnings
2025-12-04T08:53:09.5674620Z  * [new branch]                windows_libtorch_free   -> origin/windows_libtorch_free
2025-12-04T08:53:09.5674687Z  * [new branch]                xmfan-war               -> origin/xmfan-war
2025-12-04T08:53:09.5674751Z  * [new branch]                xmfan/ca_0516           -> origin/xmfan/ca_0516
2025-12-04T08:53:09.5674820Z  * [new branch]                xmfan/ca_1051b93192     -> origin/xmfan/ca_1051b93192
2025-12-04T08:53:09.5674972Z  * [new branch]                xmfan/ca_1a722f62c248391fc4a542e8851a5559aa356ae8 -> origin/xmfan/ca_1a722f62c248391fc4a542e8851a5559aa356ae8
2025-12-04T08:53:09.5675043Z  * [new branch]                xmfan/ca_5a2be192d1     -> origin/xmfan/ca_5a2be192d1
2025-12-04T08:53:09.5675111Z  * [new branch]                xmfan/ca_9d59b516e9     -> origin/xmfan/ca_9d59b516e9
2025-12-04T08:53:09.5675176Z  * [new branch]                xmfan/ca_apr8           -> origin/xmfan/ca_apr8
2025-12-04T08:53:09.5675238Z  * [new branch]                xmfan/ca_base           -> origin/xmfan/ca_base
2025-12-04T08:53:09.5675306Z  * [new branch]                xmfan/ca_dynamic        -> origin/xmfan/ca_dynamic
2025-12-04T08:53:09.5675402Z  * [new branch]                xmfan/ca_fix_dyn        -> origin/xmfan/ca_fix_dyn
2025-12-04T08:53:09.5675503Z  * [new branch]                xmfan/ca_fix_lowering   -> origin/xmfan/ca_fix_lowering
2025-12-04T08:53:09.5675579Z  * [new branch]                xmfan/ca_fix_polyfills  -> origin/xmfan/ca_fix_polyfills
2025-12-04T08:53:09.5675642Z  * [new branch]                xmfan/ca_jan3           -> origin/xmfan/ca_jan3
2025-12-04T08:53:09.5675706Z  * [new branch]                xmfan/ca_jun18          -> origin/xmfan/ca_jun18
2025-12-04T08:53:09.5675772Z  * [new branch]                xmfan/ca_jun24          -> origin/xmfan/ca_jun24
2025-12-04T08:53:09.5675839Z  * [new branch]                xmfan/ca_nested         -> origin/xmfan/ca_nested
2025-12-04T08:53:09.5675906Z  * [new branch]                xmfan/ca_overhead       -> origin/xmfan/ca_overhead
2025-12-04T08:53:09.5675999Z  * [new branch]                xmfan/ca_overhead_0eba7e5451 -> origin/xmfan/ca_overhead_0eba7e5451
2025-12-04T08:53:09.5676069Z  * [new branch]                xmfan/cacu_jun18        -> origin/xmfan/cacu_jun18
2025-12-04T08:53:09.5676137Z  * [new branch]                xmfan/cacu_jun19        -> origin/xmfan/cacu_jun19
2025-12-04T08:53:09.5676203Z  * [new branch]                xmfan/cacu_jun4         -> origin/xmfan/cacu_jun4
2025-12-04T08:53:09.5676285Z  * [new branch]                xmfan/disable_duck_shape -> origin/xmfan/disable_duck_shape
2025-12-04T08:53:09.5676383Z  * [new branch]                xmfan/fca_cpp_node_passthrough -> origin/xmfan/fca_cpp_node_passthrough
2025-12-04T08:53:09.5676540Z  * [new branch]                xmfan/post_3945954741e2d37023c5d6954f9483008e0892f9 -> origin/xmfan/post_3945954741e2d37023c5d6954f9483008e0892f9
2025-12-04T08:53:09.5676689Z  * [new branch]                xmfan/pre_3945954741e2d37023c5d6954f9483008e0892f9 -> origin/xmfan/pre_3945954741e2d37023c5d6954f9483008e0892f9
2025-12-04T08:53:09.5676760Z  * [new branch]                xmfan/single_step       -> origin/xmfan/single_step
2025-12-04T08:53:09.5676826Z  * [new branch]                xmfan/sth_0829          -> origin/xmfan/sth_0829
2025-12-04T08:53:09.5676889Z  * [new branch]                xmfan/test              -> origin/xmfan/test
2025-12-04T08:53:09.5676979Z  * [new branch]                yguo/debug-0226-constexpr -> origin/yguo/debug-0226-constexpr
2025-12-04T08:53:09.5677056Z  * [new branch]                yguo/new_latest_changes -> origin/yguo/new_latest_changes
2025-12-04T08:53:09.5677151Z  * [new branch]                yguo/patch_constexpr_changes -> origin/yguo/patch_constexpr_changes
2025-12-04T08:53:09.5677220Z  * [new branch]                yiming/bootcamp         -> origin/yiming/bootcamp
2025-12-04T08:53:09.5677321Z  * [new branch]                yiming/run_with_start_end_rng_hop -> origin/yiming/run_with_start_end_rng_hop
2025-12-04T08:53:09.5677385Z  * [new branch]                yolo-llama3             -> origin/yolo-llama3
2025-12-04T08:53:09.5677459Z  * [new branch]                zainr/canary-test       -> origin/zainr/canary-test
2025-12-04T08:53:09.5677547Z  * [new branch]                zainr/cleanup-gh-runners -> origin/zainr/cleanup-gh-runners
2025-12-04T08:53:09.5677629Z  * [new branch]                zainr/pull-migration-c  -> origin/zainr/pull-migration-c
2025-12-04T08:53:09.5677692Z  * [new branch]                zainr/test2             -> origin/zainr/test2
2025-12-04T08:53:09.5677764Z  * [new branch]                zasdfgbnm-patch-3       -> origin/zasdfgbnm-patch-3
2025-12-04T08:53:09.5677825Z  * [new branch]                zb2p                    -> origin/zb2p
2025-12-04T08:53:09.5677911Z  * [new branch]                zeros-and-scatter-part2 -> origin/zeros-and-scatter-part2
2025-12-04T08:53:09.5677998Z  * [new branch]                zhxchen17/ci/vllm_lora_oom -> origin/zhxchen17/ci/vllm_lora_oom
2025-12-04T08:53:09.5678100Z  * [new branch]                zhxchen17/ci/vllm_multimodal_oom -> origin/zhxchen17/ci/vllm_multimodal_oom
2025-12-04T08:53:09.5678206Z  * [new branch]                zhxchen17/ci/vllm_pin   -> origin/zhxchen17/ci/vllm_pin
2025-12-04T08:53:09.5678351Z  * [new branch]                zhxchen17/dynamo/unsafe_drop_all_guards -> origin/zhxchen17/dynamo/unsafe_drop_all_guards
2025-12-04T08:53:09.5678450Z  * [new branch]                zhxchen17/export/call_override -> origin/zhxchen17/export/call_override
2025-12-04T08:53:09.5678538Z  * [new branch]                zhxchen17/export/codemod1 -> origin/zhxchen17/export/codemod1
2025-12-04T08:53:09.5678627Z  * [new branch]                zhxchen17/export/ctx_return -> origin/zhxchen17/export/ctx_return
2025-12-04T08:53:09.5678758Z  * [new branch]                zhxchen17/export/disable_side_effect_warn -> origin/zhxchen17/export/disable_side_effect_warn
2025-12-04T08:53:09.5678856Z  * [new branch]                zhxchen17/export/pytree_check -> origin/zhxchen17/export/pytree_check
2025-12-04T08:53:09.5678942Z  * [new branch]                zhxchen17/precompile/aoti -> origin/zhxchen17/precompile/aoti
2025-12-04T08:53:09.5679043Z  * [new branch]                zhxchen17/precompile/globals -> origin/zhxchen17/precompile/globals
2025-12-04T08:53:09.5679164Z  * [new branch]                zhxchen17/precompile/inductor_guards -> origin/zhxchen17/precompile/inductor_guards
2025-12-04T08:53:09.5679241Z  * [new branch]                zhxchen17/scratch/0     -> origin/zhxchen17/scratch/0
2025-12-04T08:53:09.5679350Z  * [new branch]                zhxchen17/torch_export_api_update -> origin/zhxchen17/torch_export_api_update
2025-12-04T08:53:09.5679427Z  * [new branch]                zhxhcen17/moodycamel    -> origin/zhxhcen17/moodycamel
2025-12-04T08:53:09.5679505Z  * [new branch]                zxiiro/build-times      -> origin/zxiiro/build-times
2025-12-04T08:53:09.5679583Z  * [new branch]                zxiiro/c7i.2xlarge      -> origin/zxiiro/c7i.2xlarge
2025-12-04T08:53:09.5679665Z  * [new branch]                zxiiro/c7i.2xlarge.h100 -> origin/zxiiro/c7i.2xlarge.h100
2025-12-04T08:53:09.5679733Z  * [new branch]                zxiiro/main             -> origin/zxiiro/main
2025-12-04T08:53:09.5679806Z  * [new branch]                zxiiro/risc64           -> origin/zxiiro/risc64
2025-12-04T08:53:09.5679901Z  * [new branch]                zxiiro/test-multicloud-arc -> origin/zxiiro/test-multicloud-arc
2025-12-04T08:53:09.5679963Z  * [new tag]                   ciflow/dynamo/169525    -> ciflow/dynamo/169525
2025-12-04T08:53:09.5680036Z  t [tag update]                ciflow/inductor/167647  -> ciflow/inductor/167647
2025-12-04T08:53:09.5680107Z  t [tag update]                ciflow/inductor/168266  -> ciflow/inductor/168266
2025-12-04T08:53:09.5680183Z  t [tag update]                ciflow/inductor/169535  -> ciflow/inductor/169535
2025-12-04T08:53:09.5680243Z  * [new tag]                   ciflow/trunk/165728     -> ciflow/trunk/165728
2025-12-04T08:53:09.5680304Z  * [new tag]                   ciflow/trunk/169048     -> ciflow/trunk/169048
2025-12-04T08:53:09.5680366Z  * [new tag]                   ciflow/trunk/169125     -> ciflow/trunk/169125
2025-12-04T08:53:09.5680456Z  * [new tag]                   ciflow/trunk/169555     -> ciflow/trunk/169555
2025-12-04T08:53:09.5680517Z  * [new tag]                   ciflow/xpu/169555       -> ciflow/xpu/169555
2025-12-04T08:53:09.7546291Z [command]/usr/bin/git rev-parse --verify --quiet ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32^{object}
2025-12-04T08:53:09.7675291Z ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T08:53:09.7680200Z ##[endgroup]
2025-12-04T08:53:09.7680616Z ##[group]Determining the checkout info
2025-12-04T08:53:09.7681392Z ##[endgroup]
2025-12-04T08:53:09.7686290Z [command]/usr/bin/git sparse-checkout disable
2025-12-04T08:53:09.7779390Z [command]/usr/bin/git config --local --unset-all extensions.worktreeConfig
2025-12-04T08:53:09.7803386Z ##[group]Checking out the ref
2025-12-04T08:53:09.7805102Z [command]/usr/bin/git checkout --progress --force ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T08:53:09.8077430Z HEAD is now at ffd9b0fb4355 Resolve collective autotuning test failure on arm (#168919)
2025-12-04T08:53:09.8083784Z ##[endgroup]
2025-12-04T08:53:09.8084074Z ##[group]Setting up auth for fetching submodules
2025-12-04T08:53:09.8089043Z [command]/usr/bin/git config --global http.https://github.com/.extraheader AUTHORIZATION: basic ***
2025-12-04T08:53:09.8120700Z [command]/usr/bin/git config --global --unset-all url.https://github.com/.insteadOf
2025-12-04T08:53:09.8146780Z [command]/usr/bin/git config --global --add url.https://github.com/.insteadOf git@github.com:
2025-12-04T08:53:09.8171007Z [command]/usr/bin/git config --global --add url.https://github.com/.insteadOf org-21003710@github.com:
2025-12-04T08:53:09.8191909Z ##[endgroup]
2025-12-04T08:53:09.8192072Z ##[group]Fetching submodules
2025-12-04T08:53:09.8193751Z [command]/usr/bin/git submodule sync --recursive
2025-12-04T08:53:09.8429248Z Synchronizing submodule url for 'android/libs/fbjni'
2025-12-04T08:53:09.8446679Z Synchronizing submodule url for 'third_party/FP16'
2025-12-04T08:53:09.8458896Z Synchronizing submodule url for 'third_party/FXdiv'
2025-12-04T08:53:09.8475080Z Synchronizing submodule url for 'third_party/NNPACK'
2025-12-04T08:53:09.8487223Z Synchronizing submodule url for 'third_party/NVTX'
2025-12-04T08:53:09.8498505Z Synchronizing submodule url for 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:09.8510256Z Synchronizing submodule url for 'third_party/XNNPACK'
2025-12-04T08:53:09.8525544Z Synchronizing submodule url for 'third_party/aiter'
2025-12-04T08:53:09.8537587Z Synchronizing submodule url for 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:09.8572905Z Synchronizing submodule url for 'third_party/benchmark'
2025-12-04T08:53:09.8589726Z Synchronizing submodule url for 'third_party/composable_kernel'
2025-12-04T08:53:09.8608193Z Synchronizing submodule url for 'third_party/cpp-httplib'
2025-12-04T08:53:09.8619220Z Synchronizing submodule url for 'third_party/cpuinfo'
2025-12-04T08:53:09.8629199Z Synchronizing submodule url for 'third_party/cudnn_frontend'
2025-12-04T08:53:09.8638874Z Synchronizing submodule url for 'third_party/cutlass'
2025-12-04T08:53:09.8652994Z Synchronizing submodule url for 'third_party/fbgemm'
2025-12-04T08:53:09.8666011Z Synchronizing submodule url for 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:09.8676618Z Synchronizing submodule url for 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:09.8693799Z Synchronizing submodule url for 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:09.8705467Z Synchronizing submodule url for 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:09.8720377Z Synchronizing submodule url for 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:09.8731403Z Synchronizing submodule url for 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:09.8743199Z Synchronizing submodule url for 'third_party/fbgemm/external/json'
2025-12-04T08:53:09.8756452Z Synchronizing submodule url for 'third_party/flash-attention'
2025-12-04T08:53:09.8769914Z Synchronizing submodule url for 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:09.8784494Z Synchronizing submodule url for 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:09.8801076Z Synchronizing submodule url for 'third_party/flatbuffers'
2025-12-04T08:53:09.8822067Z Synchronizing submodule url for 'third_party/fmt'
2025-12-04T08:53:09.8834495Z Synchronizing submodule url for 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:09.8844631Z Synchronizing submodule url for 'third_party/gloo'
2025-12-04T08:53:09.8856340Z Synchronizing submodule url for 'third_party/googletest'
2025-12-04T08:53:09.8865848Z Synchronizing submodule url for 'third_party/ideep'
2025-12-04T08:53:09.8875156Z Synchronizing submodule url for 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:09.8889358Z Synchronizing submodule url for 'third_party/ittapi'
2025-12-04T08:53:09.8899625Z Synchronizing submodule url for 'third_party/kineto'
2025-12-04T08:53:09.8915939Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:09.8927822Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:09.8945615Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:09.8957319Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:09.8974870Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:09.8985671Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:09.8998253Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:09.9010514Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:09.9020572Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:09.9033104Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:09.9043793Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:09.9055230Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:09.9067323Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:09.9083335Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:09.9098394Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:09.9111980Z Synchronizing submodule url for 'third_party/kleidiai'
2025-12-04T08:53:09.9123868Z Synchronizing submodule url for 'third_party/mimalloc'
2025-12-04T08:53:09.9136332Z Synchronizing submodule url for 'third_party/nlohmann'
2025-12-04T08:53:09.9147415Z Synchronizing submodule url for 'third_party/onnx'
2025-12-04T08:53:09.9167322Z Synchronizing submodule url for 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:09.9181323Z Synchronizing submodule url for 'third_party/opentelemetry-cpp'
2025-12-04T08:53:09.9197846Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:09.9208208Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:09.9218981Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:09.9235422Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:09.9247081Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:09.9258923Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:09.9270763Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:09.9283340Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:09.9293331Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:09.9306286Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:09.9325929Z Synchronizing submodule url for 'third_party/pocketfft'
2025-12-04T08:53:09.9336021Z Synchronizing submodule url for 'third_party/protobuf'
2025-12-04T08:53:09.9348270Z Synchronizing submodule url for 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:09.9359076Z Synchronizing submodule url for 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:09.9371874Z Synchronizing submodule url for 'third_party/psimd'
2025-12-04T08:53:09.9381677Z Synchronizing submodule url for 'third_party/pthreadpool'
2025-12-04T08:53:09.9393201Z Synchronizing submodule url for 'third_party/pybind11'
2025-12-04T08:53:09.9403594Z Synchronizing submodule url for 'third_party/python-peachpy'
2025-12-04T08:53:09.9414500Z Synchronizing submodule url for 'third_party/sleef'
2025-12-04T08:53:09.9425143Z Synchronizing submodule url for 'third_party/tensorpipe'
2025-12-04T08:53:09.9437187Z Synchronizing submodule url for 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:09.9447342Z Synchronizing submodule url for 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:09.9457730Z Synchronizing submodule url for 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:09.9468584Z Synchronizing submodule url for 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:09.9486223Z Synchronizing submodule url for 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:09.9512188Z [command]/usr/bin/git -c protocol.version=2 submodule update --init --force --recursive
2025-12-04T08:53:09.9759552Z Submodule path 'android/libs/fbjni': checked out '7e1e1fe3858c63c251c637ae41a20de425dde96f'
2025-12-04T08:53:09.9832333Z Submodule path 'third_party/FP16': checked out '4dfe081cf6bcd15db339cf2680b9281b8451eeb3'
2025-12-04T08:53:09.9884337Z Submodule path 'third_party/FXdiv': checked out 'b408327ac2a15ec3e43352421954f5b1967701d1'
2025-12-04T08:53:09.9997151Z Submodule path 'third_party/NNPACK': checked out 'c07e3a0400713d546e0dea2d5466dd22ea389c73'
2025-12-04T08:53:10.0063433Z Submodule path 'third_party/NVTX': checked out '3ebbc93ded7285963bff932c678fa367eb393ba6'
2025-12-04T08:53:10.0115959Z Submodule path 'third_party/VulkanMemoryAllocator': checked out '1d8f600fd424278486eade7ed3e877c99f0846b1'
2025-12-04T08:53:10.5005965Z Submodule path 'third_party/XNNPACK': checked out '51a0103656eff6fc9bfd39a4597923c4b542c883'
2025-12-04T08:53:10.5175551Z Submodule path 'third_party/aiter': checked out '01aae101b9e5e94d6c16a9514c9fb8df99c93150'
2025-12-04T08:53:10.5397877Z Submodule path 'third_party/aiter/3rdparty/composable_kernel': checked out 'cffe8fa2a442ac8e80dd236a1a5d24fe3d7e0cbf'
2025-12-04T08:53:10.5526101Z Submodule path 'third_party/benchmark': checked out '299e5928955cc62af9968370293b916f5130916f'
2025-12-04T08:53:10.5710885Z Submodule path 'third_party/composable_kernel': checked out '7fe50dc3da2069d6645d9deb8c017a876472a977'
2025-12-04T08:53:10.5782808Z Submodule path 'third_party/cpp-httplib': checked out '89c932f313c6437c38f2982869beacc89c2f2246'
2025-12-04T08:53:10.6424850Z Submodule path 'third_party/cpuinfo': checked out 'f858c30bcb16f8effd5ff46996f0514539e17abc'
2025-12-04T08:53:10.6505968Z Submodule path 'third_party/cudnn_frontend': checked out '0b1577c8c83401237d601d0d0db5210506705396'
2025-12-04T08:53:10.6643156Z Submodule path 'third_party/cutlass': checked out 'f88806b1e31dfa579842638740216dd41fc6c588'
2025-12-04T08:53:10.7357894Z Submodule path 'third_party/fbgemm': checked out 'c0b988d39a9e47c794d699f29930ed4d7c7e13a4'
2025-12-04T08:53:10.7670014Z Submodule path 'third_party/fbgemm/external/asmjit': checked out 'a3199e8857792cd10b7589ff5d58343d2c9008ea'
2025-12-04T08:53:10.9510132Z Submodule path 'third_party/fbgemm/external/composable_kernel': checked out '7fe50dc3da2069d6645d9deb8c017a876472a977'
2025-12-04T08:53:11.0189203Z Submodule path 'third_party/fbgemm/external/cpuinfo': checked out '6543fec09b2f04ac4a666882998b534afc9c1349'
2025-12-04T08:53:11.4709132Z Submodule path 'third_party/fbgemm/external/cutlass': checked out '98125ce499b0fdf7ffbe0e3052f5b8709f4840f8'
2025-12-04T08:53:11.4927474Z Submodule path 'third_party/fbgemm/external/googletest': checked out '52eb8108c5bdec04579160ae17225d66034bd723'
2025-12-04T08:53:11.5010968Z Submodule path 'third_party/fbgemm/external/hipify_torch': checked out '63b6a7b541fa7f08f8475ca7d74054db36ff2691'
2025-12-04T08:53:11.5570078Z Submodule path 'third_party/fbgemm/external/json': checked out '9cca280a4d0ccf0c08f47a99aa71d1b0e52f8d03'
2025-12-04T08:53:11.5671169Z Submodule path 'third_party/flash-attention': checked out '979702c87a8713a8e0a5e9fee122b90d2ef13be5'
2025-12-04T08:53:11.5865573Z Submodule path 'third_party/flash-attention/csrc/composable_kernel': checked out '888317e698e9803c62bd38568abc9e05d7709f33'
2025-12-04T08:53:11.5990554Z Submodule path 'third_party/flash-attention/csrc/cutlass': checked out 'c506e16788cb08416a4a57e11a9067beeee29420'
2025-12-04T08:53:11.6095257Z Submodule path 'third_party/flatbuffers': checked out 'a2cd1ea3b6d3fee220106b5fed3f7ce8da9eb757'
2025-12-04T08:53:11.6237604Z Submodule path 'third_party/fmt': checked out '407c905e45ad75fc29bf0f9bb7c5c2fd3475976f'
2025-12-04T08:53:11.6440262Z Submodule path 'third_party/gemmlowp/gemmlowp': checked out '3fb5c176c17c765a3492cd2f0321b0dab712f350'
2025-12-04T08:53:11.6570130Z Submodule path 'third_party/gloo': checked out '54cbae0d3a67fa890b4c3d9ee162b7860315e341'
2025-12-04T08:53:11.6762391Z Submodule path 'third_party/googletest': checked out '52eb8108c5bdec04579160ae17225d66034bd723'
2025-12-04T08:53:11.6832214Z Submodule path 'third_party/ideep': checked out '719d8e6cd7f7a0e01b155657526d693acf97c2b3'
2025-12-04T08:53:12.0966117Z Submodule path 'third_party/ideep/mkl-dnn': checked out '8d263e693366ef8db40acc569cc7d8edf644556d'
2025-12-04T08:53:12.1069953Z Submodule path 'third_party/ittapi': checked out 'dec1d23ca65ab069d225dfe40dea14f455170959'
2025-12-04T08:53:12.1161561Z Submodule path 'third_party/kineto': checked out '31f85df8fbd89c188f14ef10f1ec65379786b943'
2025-12-04T08:53:12.1268539Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog': checked out 'd2ffe0a4e3acace628db49974246b66fc3e85fb1'
2025-12-04T08:53:12.1345694Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM': checked out 'ffde4e54bc7249a6039a5e6b45b395141e1217f9'
2025-12-04T08:53:12.1421418Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr': checked out '871ed52d350214a034f6ef8a3b8f51c5ce1bd400'
2025-12-04T08:53:12.1486664Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt': checked out 'cd4af11efc9c622896a3e4cb599fa28668ca3d05'
2025-12-04T08:53:12.1558286Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags': checked out 'e171aa2d15ed9eb17054558e0b3a6a413bb01067'
2025-12-04T08:53:12.1609505Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc': checked out '8411df715cf522606e3b1aca386ddfc0b63d34b4'
2025-12-04T08:53:12.1667155Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog': checked out 'b33e3bad4c46c8a6345525fd822af355e5ef9446'
2025-12-04T08:53:12.1725828Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest': checked out '52eb8108c5bdec04579160ae17225d66034bd723'
2025-12-04T08:53:12.1816528Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/json': checked out '4f8fba14066156b73f1189a2b8bd568bde5284c5'
2025-12-04T08:53:12.1876665Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs': checked out 'f68a2fa8ea36c783bdd760371411fcb495aa3150'
2025-12-04T08:53:12.1933082Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp': checked out 'b1234816facfdda29845c46696a02998a4af115a'
2025-12-04T08:53:12.2018655Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb': checked out 'd7ba35bbb649209c66e582d5a0244ba988a15159'
2025-12-04T08:53:12.2081616Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest': checked out 'e2239ee6043f73722e7aa812a459f54a28552929'
2025-12-04T08:53:12.2143349Z Submodule path 'third_party/kineto/libkineto/third_party/fmt': checked out '40626af88bd7df9a5fb80be7b25ac85b122d6c21'
2025-12-04T08:53:12.2206638Z Submodule path 'third_party/kineto/libkineto/third_party/googletest': checked out '52eb8108c5bdec04579160ae17225d66034bd723'
2025-12-04T08:53:12.2279257Z Submodule path 'third_party/kleidiai': checked out 'd7770c89632329a9914ef1a90289917597639cbe'
2025-12-04T08:53:12.2359756Z Submodule path 'third_party/mimalloc': checked out 'fbd8b99c2b828428947d70fdc046bb55609be93e'
2025-12-04T08:53:12.2468592Z Submodule path 'third_party/nlohmann': checked out '55f93686c01528224f448c19128836e7df245f72'
2025-12-04T08:53:12.4154364Z Submodule path 'third_party/onnx': checked out 'e709452ef2bbc1d113faf678c24e6d3467696e83'
2025-12-04T08:53:12.4347711Z Submodule path 'third_party/onnx/third_party/pybind11': checked out 'a2e59f0e7065404b44dfe92a28aca47ba1378dc4'
2025-12-04T08:53:12.4462035Z Submodule path 'third_party/opentelemetry-cpp': checked out 'a799f4aed9c94b765dcdaabaeab7d5e7e2310878'
2025-12-04T08:53:12.4551483Z Submodule path 'third_party/opentelemetry-cpp/third_party/benchmark': checked out 'd572f4777349d43653b21d6c2fc63020ab326db2'
2025-12-04T08:53:12.4617649Z Submodule path 'third_party/opentelemetry-cpp/third_party/googletest': checked out 'b796f7d44681514f58a683a3a71ff17c94edb0c1'
2025-12-04T08:53:12.4685510Z Submodule path 'third_party/opentelemetry-cpp/third_party/ms-gsl': checked out '6f4529395c5b7c2d661812257cd6780c67e54afa'
2025-12-04T08:53:12.4786570Z Submodule path 'third_party/opentelemetry-cpp/third_party/nlohmann-json': checked out 'bc889afb4c5bf1c0d8ee29ef35eaaf4c8bef8a5d'
2025-12-04T08:53:12.4838548Z Submodule path 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto': checked out '4ca4f0335c63cda7ab31ea7ed70d6553aee14dce'
2025-12-04T08:53:12.4887606Z Submodule path 'third_party/opentelemetry-cpp/third_party/opentracing-cpp': checked out '06b57f48ded1fa3bdd3d4346f6ef29e40e08eaf5'
2025-12-04T08:53:12.4947869Z Submodule path 'third_party/opentelemetry-cpp/third_party/prometheus-cpp': checked out 'c9ffcdda9086ffd9e1283ea7a0276d831f3c8a8d'
2025-12-04T08:53:12.5023030Z Submodule path 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb': checked out 'eefb26f82b233268fc98577d265352720d477ba4'
2025-12-04T08:53:12.5110564Z Submodule path 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest': checked out 'e2239ee6043f73722e7aa812a459f54a28552929'
2025-12-04T08:53:12.5271551Z Submodule path 'third_party/opentelemetry-cpp/tools/vcpkg': checked out '8eb57355a4ffb410a2e94c07b4dca2dffbee8e50'
2025-12-04T08:53:12.5343379Z Submodule path 'third_party/pocketfft': checked out '0fa0ef591e38c2758e3184c6c23e497b9f732ffa'
2025-12-04T08:53:12.6702305Z Submodule path 'third_party/protobuf': checked out 'd1eca4e4b421cd2997495c4b4e65cea6be4e9b8a'
2025-12-04T08:53:12.6797724Z Submodule path 'third_party/protobuf/third_party/benchmark': checked out '5b7683f49e1e9223cf9927b24f6fd3d6bd82e3f8'
2025-12-04T08:53:12.7017923Z Submodule path 'third_party/protobuf/third_party/googletest': checked out '5ec7f0c4a113e2f18ac2c6cc7df51ad6afc24081'
2025-12-04T08:53:12.7096900Z Submodule path 'third_party/psimd': checked out '072586a71b55b7f8c584153d223e95687148a900'
2025-12-04T08:53:12.7188242Z Submodule path 'third_party/pthreadpool': checked out '4fe0e1e183925bf8cfa6aae24237e724a96479b8'
2025-12-04T08:53:12.7382626Z Submodule path 'third_party/pybind11': checked out 'f5fbe867d2d26e4a0a9177a51f6e568868ad3dc8'
2025-12-04T08:53:12.7613757Z Submodule path 'third_party/python-peachpy': checked out 'f45429b087dd7d5bc78bb40dc7cf06425c252d67'
2025-12-04T08:53:12.7874137Z Submodule path 'third_party/sleef': checked out '5a1d179df9cf652951b59010a2d2075372d67f68'
2025-12-04T08:53:12.7984624Z Submodule path 'third_party/tensorpipe': checked out '2b4cd91092d335a697416b2a3cb398283246849d'
2025-12-04T08:53:12.8169032Z Submodule path 'third_party/tensorpipe/third_party/googletest': checked out 'aee0f9d9b5b87796ee8a0ab26b7587ec30e8858e'
2025-12-04T08:53:12.8247856Z Submodule path 'third_party/tensorpipe/third_party/libnop': checked out '910b55815be16109f04f4180e9adee14fb4ce281'
2025-12-04T08:53:12.8538769Z Submodule path 'third_party/tensorpipe/third_party/libuv': checked out '5152db2cbfeb5582e9c27c5ea1dba2cd9e10759b'
2025-12-04T08:53:12.8683999Z Submodule path 'third_party/tensorpipe/third_party/pybind11': checked out 'a23996fce38ff6ccfbcdc09f1e63f2c4be5ea2ef'
2025-12-04T08:53:12.8757829Z Submodule path 'third_party/tensorpipe/third_party/pybind11/tools/clang': checked out '6a00cbc4a9b8e68b71caf7f774b3f9c753ae84d5'
2025-12-04T08:53:12.8798090Z [command]/usr/bin/git submodule foreach --recursive git config --local gc.auto 0
2025-12-04T08:53:12.9006151Z Entering 'android/libs/fbjni'
2025-12-04T08:53:12.9028022Z Entering 'third_party/FP16'
2025-12-04T08:53:12.9049471Z Entering 'third_party/FXdiv'
2025-12-04T08:53:12.9072726Z Entering 'third_party/NNPACK'
2025-12-04T08:53:12.9094397Z Entering 'third_party/NVTX'
2025-12-04T08:53:12.9116751Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:12.9143647Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:12.9173279Z Entering 'third_party/aiter'
2025-12-04T08:53:12.9196159Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:12.9226306Z Entering 'third_party/benchmark'
2025-12-04T08:53:12.9247402Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:12.9274531Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:12.9296932Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:12.9322846Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:12.9343765Z Entering 'third_party/cutlass'
2025-12-04T08:53:12.9366943Z Entering 'third_party/fbgemm'
2025-12-04T08:53:12.9386985Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:12.9406890Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:12.9431081Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:12.9450734Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:12.9474710Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:12.9494523Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:12.9518336Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:12.9540994Z Entering 'third_party/flash-attention'
2025-12-04T08:53:12.9561548Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:12.9583657Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:12.9608780Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:12.9630063Z Entering 'third_party/fmt'
2025-12-04T08:53:12.9649745Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:12.9669057Z Entering 'third_party/gloo'
2025-12-04T08:53:12.9693790Z Entering 'third_party/googletest'
2025-12-04T08:53:12.9724554Z Entering 'third_party/ideep'
2025-12-04T08:53:12.9746625Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:12.9772975Z Entering 'third_party/ittapi'
2025-12-04T08:53:12.9802362Z Entering 'third_party/kineto'
2025-12-04T08:53:12.9825658Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:12.9846448Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:12.9868254Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:12.9890175Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:12.9910335Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:12.9929765Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:12.9951118Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:12.9974559Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:12.9995147Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:13.0027231Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:13.0041860Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:13.0060285Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:13.0079117Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:13.0103219Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:13.0139460Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:13.0161222Z Entering 'third_party/kleidiai'
2025-12-04T08:53:13.0181901Z Entering 'third_party/mimalloc'
2025-12-04T08:53:13.0205491Z Entering 'third_party/nlohmann'
2025-12-04T08:53:13.0227351Z Entering 'third_party/onnx'
2025-12-04T08:53:13.0258471Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:13.0285481Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:13.0307250Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:13.0325843Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:13.0344971Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:13.0363669Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:13.0388399Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:13.0407886Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:13.0427942Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:13.0446832Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:13.0484234Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:13.0508276Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:13.0544275Z Entering 'third_party/pocketfft'
2025-12-04T08:53:13.0564065Z Entering 'third_party/protobuf'
2025-12-04T08:53:13.0585625Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:13.0612724Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:13.0639031Z Entering 'third_party/psimd'
2025-12-04T08:53:13.0666580Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:13.0698185Z Entering 'third_party/pybind11'
2025-12-04T08:53:13.0717462Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:13.0737323Z Entering 'third_party/sleef'
2025-12-04T08:53:13.0756209Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:13.0775714Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:13.0796023Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:13.0815904Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:13.0835028Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:13.0855403Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:13.0890114Z ##[endgroup]
2025-12-04T08:53:13.0890342Z ##[group]Persisting credentials for submodules
2025-12-04T08:53:13.0898260Z [command]/usr/bin/git submodule foreach --recursive sh -c "git config --local --name-only --get-regexp 'url\.https\:\/\/github\.com\/\.insteadOf' && git config --local --unset-all 'url.https://github.com/.insteadOf' || :"
2025-12-04T08:53:13.1072779Z Entering 'android/libs/fbjni'
2025-12-04T08:53:13.1095142Z Entering 'third_party/FP16'
2025-12-04T08:53:13.1121582Z Entering 'third_party/FXdiv'
2025-12-04T08:53:13.1142112Z Entering 'third_party/NNPACK'
2025-12-04T08:53:13.1163862Z Entering 'third_party/NVTX'
2025-12-04T08:53:13.1190624Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:13.1214846Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:13.1247937Z Entering 'third_party/aiter'
2025-12-04T08:53:13.1273238Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:13.1298546Z Entering 'third_party/benchmark'
2025-12-04T08:53:13.1330272Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:13.1355866Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:13.1376468Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:13.1399804Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:13.1424150Z Entering 'third_party/cutlass'
2025-12-04T08:53:13.1450961Z Entering 'third_party/fbgemm'
2025-12-04T08:53:13.1483964Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:13.1508326Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:13.1535772Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:13.1558111Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:13.1582869Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:13.1605364Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:13.1632378Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:13.1657040Z Entering 'third_party/flash-attention'
2025-12-04T08:53:13.1680720Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:13.1706843Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:13.1736986Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:13.1762386Z Entering 'third_party/fmt'
2025-12-04T08:53:13.1785149Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:13.1806080Z Entering 'third_party/gloo'
2025-12-04T08:53:13.1830377Z Entering 'third_party/googletest'
2025-12-04T08:53:13.1854773Z Entering 'third_party/ideep'
2025-12-04T08:53:13.1882251Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:13.1908352Z Entering 'third_party/ittapi'
2025-12-04T08:53:13.1933420Z Entering 'third_party/kineto'
2025-12-04T08:53:13.1955907Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:13.1979094Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:13.2002499Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:13.2022700Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:13.2043741Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:13.2064962Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:13.2089480Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:13.2111184Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:13.2132549Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:13.2156610Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:13.2179298Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:13.2200397Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:13.2222117Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:13.2253250Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:13.2273192Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:13.2293918Z Entering 'third_party/kleidiai'
2025-12-04T08:53:13.2315723Z Entering 'third_party/mimalloc'
2025-12-04T08:53:13.2338839Z Entering 'third_party/nlohmann'
2025-12-04T08:53:13.2362778Z Entering 'third_party/onnx'
2025-12-04T08:53:13.2392994Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:13.2419258Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:13.2445286Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:13.2472854Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:13.2496604Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:13.2518105Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:13.2541824Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:13.2563087Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:13.2585319Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:13.2606939Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:13.2632860Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:13.2660754Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:13.2692682Z Entering 'third_party/pocketfft'
2025-12-04T08:53:13.2716315Z Entering 'third_party/protobuf'
2025-12-04T08:53:13.2741525Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:13.2772984Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:13.2794506Z Entering 'third_party/psimd'
2025-12-04T08:53:13.2817896Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:13.2844736Z Entering 'third_party/pybind11'
2025-12-04T08:53:13.2868025Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:13.2888265Z Entering 'third_party/sleef'
2025-12-04T08:53:13.2910740Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:13.2931634Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:13.2957645Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:13.2980691Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:13.3004687Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:13.3029750Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:13.3068082Z [command]/usr/bin/git submodule foreach --recursive sh -c "git config --local 'http.https://github.com/.extraheader' 'AUTHORIZATION: basic ***' && git config --local --show-origin --name-only --get-regexp remote.origin.url"
2025-12-04T08:53:13.3238814Z Entering 'android/libs/fbjni'
2025-12-04T08:53:13.3258325Z file:/home/runner/_work/pytorch/pytorch/.git/modules/android/libs/fbjni/config	remote.origin.url
2025-12-04T08:53:13.3268918Z Entering 'third_party/FP16'
2025-12-04T08:53:13.3292200Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FP16/config	remote.origin.url
2025-12-04T08:53:13.3302508Z Entering 'third_party/FXdiv'
2025-12-04T08:53:13.3325457Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FXdiv/config	remote.origin.url
2025-12-04T08:53:13.3337591Z Entering 'third_party/NNPACK'
2025-12-04T08:53:13.3358601Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK/config	remote.origin.url
2025-12-04T08:53:13.3369147Z Entering 'third_party/NVTX'
2025-12-04T08:53:13.3389151Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NVTX/config	remote.origin.url
2025-12-04T08:53:13.3398861Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:13.3421167Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/VulkanMemoryAllocator/config	remote.origin.url
2025-12-04T08:53:13.3431499Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:13.3452853Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/XNNPACK/config	remote.origin.url
2025-12-04T08:53:13.3474244Z Entering 'third_party/aiter'
2025-12-04T08:53:13.3502914Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/config	remote.origin.url
2025-12-04T08:53:13.3515218Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:13.3535202Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/modules/3rdparty/composable_kernel/config	remote.origin.url
2025-12-04T08:53:13.3549594Z Entering 'third_party/benchmark'
2025-12-04T08:53:13.3570252Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T08:53:13.3580723Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:13.3599624Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/composable_kernel/config	remote.origin.url
2025-12-04T08:53:13.3613638Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:13.3633282Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpp-httplib/config	remote.origin.url
2025-12-04T08:53:13.3653629Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:13.3685880Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpuinfo/config	remote.origin.url
2025-12-04T08:53:13.3697769Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:13.3727806Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cudnn_frontend/config	remote.origin.url
2025-12-04T08:53:13.3747899Z Entering 'third_party/cutlass'
2025-12-04T08:53:13.3775684Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cutlass/config	remote.origin.url
2025-12-04T08:53:13.3797584Z Entering 'third_party/fbgemm'
2025-12-04T08:53:13.3823417Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/config	remote.origin.url
2025-12-04T08:53:13.3833611Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:13.3861312Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/asmjit/config	remote.origin.url
2025-12-04T08:53:13.3872404Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:13.3915318Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/composable_kernel/config	remote.origin.url
2025-12-04T08:53:13.3934461Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:13.3958231Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cpuinfo/config	remote.origin.url
2025-12-04T08:53:13.3972006Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:13.3997477Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cutlass/config	remote.origin.url
2025-12-04T08:53:13.4011308Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:13.4035548Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/googletest/config	remote.origin.url
2025-12-04T08:53:13.4047521Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:13.4068961Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/hipify_torch/config	remote.origin.url
2025-12-04T08:53:13.4077908Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:13.4101989Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/json/config	remote.origin.url
2025-12-04T08:53:13.4114803Z Entering 'third_party/flash-attention'
2025-12-04T08:53:13.4133781Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/config	remote.origin.url
2025-12-04T08:53:13.4150455Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:13.4177554Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/composable_kernel/config	remote.origin.url
2025-12-04T08:53:13.4194068Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:13.4216865Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/cutlass/config	remote.origin.url
2025-12-04T08:53:13.4233539Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:13.4254756Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flatbuffers/config	remote.origin.url
2025-12-04T08:53:13.4265451Z Entering 'third_party/fmt'
2025-12-04T08:53:13.4292963Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fmt/config	remote.origin.url
2025-12-04T08:53:13.4308702Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:13.4330271Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/gemmlowp/gemmlowp/config	remote.origin.url
2025-12-04T08:53:13.4340032Z Entering 'third_party/gloo'
2025-12-04T08:53:13.4361937Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/gloo/config	remote.origin.url
2025-12-04T08:53:13.4374272Z Entering 'third_party/googletest'
2025-12-04T08:53:13.4398036Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:13.4408413Z Entering 'third_party/ideep'
2025-12-04T08:53:13.4428765Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/config	remote.origin.url
2025-12-04T08:53:13.4438481Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:13.4459116Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/modules/mkl-dnn/config	remote.origin.url
2025-12-04T08:53:13.4473550Z Entering 'third_party/ittapi'
2025-12-04T08:53:13.4494996Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ittapi/config	remote.origin.url
2025-12-04T08:53:13.4505040Z Entering 'third_party/kineto'
2025-12-04T08:53:13.4525510Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/config	remote.origin.url
2025-12-04T08:53:13.4535098Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:13.4555560Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/config	remote.origin.url
2025-12-04T08:53:13.4566157Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:13.4588162Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/DCGM/config	remote.origin.url
2025-12-04T08:53:13.4598704Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:13.4618937Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/cpr/config	remote.origin.url
2025-12-04T08:53:13.4629285Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:13.4649503Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/fmt/config	remote.origin.url
2025-12-04T08:53:13.4659282Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:13.4677956Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/config	remote.origin.url
2025-12-04T08:53:13.4687425Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:13.4709215Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/modules/doc/config	remote.origin.url
2025-12-04T08:53:13.4720020Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:13.4739539Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/glog/config	remote.origin.url
2025-12-04T08:53:13.4750591Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:13.4768033Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:13.4778624Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:13.4796475Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/json/config	remote.origin.url
2025-12-04T08:53:13.4806220Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:13.4831843Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/pfs/config	remote.origin.url
2025-12-04T08:53:13.4841734Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:13.4867718Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/config	remote.origin.url
2025-12-04T08:53:13.4880635Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:13.4902289Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/civetweb/config	remote.origin.url
2025-12-04T08:53:13.4914165Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:13.4935246Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/googletest/config	remote.origin.url
2025-12-04T08:53:13.4948396Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:13.4969981Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/fmt/config	remote.origin.url
2025-12-04T08:53:13.4978563Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:13.4998035Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:13.5009829Z Entering 'third_party/kleidiai'
2025-12-04T08:53:13.5030305Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kleidiai/config	remote.origin.url
2025-12-04T08:53:13.5047667Z Entering 'third_party/mimalloc'
2025-12-04T08:53:13.5066392Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/mimalloc/config	remote.origin.url
2025-12-04T08:53:13.5082701Z Entering 'third_party/nlohmann'
2025-12-04T08:53:13.5104181Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/nlohmann/config	remote.origin.url
2025-12-04T08:53:13.5115087Z Entering 'third_party/onnx'
2025-12-04T08:53:13.5136678Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/config	remote.origin.url
2025-12-04T08:53:13.5153078Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:13.5178773Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T08:53:13.5192211Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:13.5214498Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/config	remote.origin.url
2025-12-04T08:53:13.5225841Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:13.5246215Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T08:53:13.5255970Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:13.5278435Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:13.5289057Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:13.5313887Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/ms-gsl/config	remote.origin.url
2025-12-04T08:53:13.5324050Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:13.5343347Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/nlohmann-json/config	remote.origin.url
2025-12-04T08:53:13.5353339Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:13.5371863Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentelemetry-proto/config	remote.origin.url
2025-12-04T08:53:13.5381562Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:13.5407475Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentracing-cpp/config	remote.origin.url
2025-12-04T08:53:13.5416498Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:13.5435751Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/config	remote.origin.url
2025-12-04T08:53:13.5445133Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:13.5465158Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/civetweb/config	remote.origin.url
2025-12-04T08:53:13.5475545Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:13.5496555Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/googletest/config	remote.origin.url
2025-12-04T08:53:13.5507921Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:13.5529299Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/tools/vcpkg/config	remote.origin.url
2025-12-04T08:53:13.5547065Z Entering 'third_party/pocketfft'
2025-12-04T08:53:13.5564573Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/pocketfft/config	remote.origin.url
2025-12-04T08:53:13.5574400Z Entering 'third_party/protobuf'
2025-12-04T08:53:13.5595693Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/config	remote.origin.url
2025-12-04T08:53:13.5605525Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:13.5625956Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T08:53:13.5639113Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:13.5659708Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:13.5674118Z Entering 'third_party/psimd'
2025-12-04T08:53:13.5700347Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/psimd/config	remote.origin.url
2025-12-04T08:53:13.5710839Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:13.5731607Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/pthreadpool/config	remote.origin.url
2025-12-04T08:53:13.5742685Z Entering 'third_party/pybind11'
2025-12-04T08:53:13.5761936Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T08:53:13.5774068Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:13.5794484Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/python-peachpy/config	remote.origin.url
2025-12-04T08:53:13.5804505Z Entering 'third_party/sleef'
2025-12-04T08:53:13.5825524Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/sleef/config	remote.origin.url
2025-12-04T08:53:13.5836192Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:13.5866858Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/config	remote.origin.url
2025-12-04T08:53:13.5876843Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:13.5897318Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:13.5909996Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:13.5935051Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libnop/config	remote.origin.url
2025-12-04T08:53:13.5944242Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:13.5962486Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libuv/config	remote.origin.url
2025-12-04T08:53:13.5972409Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:13.5991395Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T08:53:13.6003588Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:13.6023015Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/modules/tools/clang/config	remote.origin.url
2025-12-04T08:53:13.6238368Z [command]/usr/bin/git submodule foreach --recursive git config --local --add 'url.https://github.com/.insteadOf' 'git@github.com:'
2025-12-04T08:53:13.6416985Z Entering 'android/libs/fbjni'
2025-12-04T08:53:13.6439475Z Entering 'third_party/FP16'
2025-12-04T08:53:13.6466327Z Entering 'third_party/FXdiv'
2025-12-04T08:53:13.6487788Z Entering 'third_party/NNPACK'
2025-12-04T08:53:13.6509242Z Entering 'third_party/NVTX'
2025-12-04T08:53:13.6531321Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:13.6552606Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:13.6577026Z Entering 'third_party/aiter'
2025-12-04T08:53:13.6596620Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:13.6621386Z Entering 'third_party/benchmark'
2025-12-04T08:53:13.6642386Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:13.6663521Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:13.6681508Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:13.6701519Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:13.6723349Z Entering 'third_party/cutlass'
2025-12-04T08:53:13.6753576Z Entering 'third_party/fbgemm'
2025-12-04T08:53:13.6773355Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:13.6799320Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:13.6821855Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:13.6843257Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:13.6867392Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:13.6887535Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:13.6913969Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:13.6944819Z Entering 'third_party/flash-attention'
2025-12-04T08:53:13.6969706Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:13.6993221Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:13.7018980Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:13.7040369Z Entering 'third_party/fmt'
2025-12-04T08:53:13.7060171Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:13.7080996Z Entering 'third_party/gloo'
2025-12-04T08:53:13.7100185Z Entering 'third_party/googletest'
2025-12-04T08:53:13.7125019Z Entering 'third_party/ideep'
2025-12-04T08:53:13.7144408Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:13.7170583Z Entering 'third_party/ittapi'
2025-12-04T08:53:13.7193070Z Entering 'third_party/kineto'
2025-12-04T08:53:13.7213430Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:13.7239372Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:13.7260170Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:13.7280353Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:13.7299109Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:13.7319445Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:13.7341890Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:13.7361414Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:13.7382944Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:13.7401869Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:13.7422181Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:13.7442246Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:13.7464954Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:13.7488884Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:13.7507773Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:13.7532615Z Entering 'third_party/kleidiai'
2025-12-04T08:53:13.7557585Z Entering 'third_party/mimalloc'
2025-12-04T08:53:13.7578196Z Entering 'third_party/nlohmann'
2025-12-04T08:53:13.7599189Z Entering 'third_party/onnx'
2025-12-04T08:53:13.7624978Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:13.7654553Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:13.7675127Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:13.7697512Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:13.7717529Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:13.7739458Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:13.7759333Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:13.7777061Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:13.7796780Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:13.7814181Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:13.7835139Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:13.7857794Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:13.7886688Z Entering 'third_party/pocketfft'
2025-12-04T08:53:13.7908549Z Entering 'third_party/protobuf'
2025-12-04T08:53:13.7929502Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:13.7947693Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:13.7970238Z Entering 'third_party/psimd'
2025-12-04T08:53:13.7988486Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:13.8008900Z Entering 'third_party/pybind11'
2025-12-04T08:53:13.8030241Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:13.8065133Z Entering 'third_party/sleef'
2025-12-04T08:53:13.8085497Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:13.8109614Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:13.8132721Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:13.8152541Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:13.8171583Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:13.8200746Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:13.8246767Z [command]/usr/bin/git submodule foreach --recursive git config --local --add 'url.https://github.com/.insteadOf' 'org-21003710@github.com:'
2025-12-04T08:53:13.8419959Z Entering 'android/libs/fbjni'
2025-12-04T08:53:13.8447369Z Entering 'third_party/FP16'
2025-12-04T08:53:13.8467471Z Entering 'third_party/FXdiv'
2025-12-04T08:53:13.8488689Z Entering 'third_party/NNPACK'
2025-12-04T08:53:13.8513285Z Entering 'third_party/NVTX'
2025-12-04T08:53:13.8537733Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:13.8557030Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:13.8581282Z Entering 'third_party/aiter'
2025-12-04T08:53:13.8600479Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:13.8631945Z Entering 'third_party/benchmark'
2025-12-04T08:53:13.8662815Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:13.8692179Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:13.8715288Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:13.8737739Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:13.8769334Z Entering 'third_party/cutlass'
2025-12-04T08:53:13.8794591Z Entering 'third_party/fbgemm'
2025-12-04T08:53:13.8820702Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:13.8848282Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:13.8872886Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:13.8896719Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:13.8920231Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:13.8944240Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:13.8962843Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:13.8984770Z Entering 'third_party/flash-attention'
2025-12-04T08:53:13.9006130Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:13.9026407Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:13.9051198Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:13.9072955Z Entering 'third_party/fmt'
2025-12-04T08:53:13.9091904Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:13.9114447Z Entering 'third_party/gloo'
2025-12-04T08:53:13.9133518Z Entering 'third_party/googletest'
2025-12-04T08:53:13.9154743Z Entering 'third_party/ideep'
2025-12-04T08:53:13.9183602Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:13.9220212Z Entering 'third_party/ittapi'
2025-12-04T08:53:13.9241162Z Entering 'third_party/kineto'
2025-12-04T08:53:13.9261691Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:13.9288080Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:13.9311088Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:13.9343653Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:13.9367867Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:13.9386621Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:13.9414273Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:13.9443876Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:13.9469150Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:13.9489305Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:13.9508774Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:13.9528300Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:13.9561615Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:13.9586839Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:13.9605116Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:13.9626404Z Entering 'third_party/kleidiai'
2025-12-04T08:53:13.9645784Z Entering 'third_party/mimalloc'
2025-12-04T08:53:13.9665159Z Entering 'third_party/nlohmann'
2025-12-04T08:53:13.9685761Z Entering 'third_party/onnx'
2025-12-04T08:53:13.9718558Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:13.9744507Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:13.9774186Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:13.9797277Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:13.9819856Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:13.9842154Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:13.9863985Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:13.9883059Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:13.9902965Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:13.9922795Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:13.9941578Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:13.9961213Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:13.9990632Z Entering 'third_party/pocketfft'
2025-12-04T08:53:14.0011166Z Entering 'third_party/protobuf'
2025-12-04T08:53:14.0032043Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:14.0053585Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:14.0077359Z Entering 'third_party/psimd'
2025-12-04T08:53:14.0099389Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:14.0120231Z Entering 'third_party/pybind11'
2025-12-04T08:53:14.0139082Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:14.0158550Z Entering 'third_party/sleef'
2025-12-04T08:53:14.0180893Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:14.0205018Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:14.0227794Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:14.0248166Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:14.0268307Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:14.0290014Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:14.0322695Z ##[endgroup]
2025-12-04T08:53:14.0458245Z [command]/usr/bin/git log -1 --format=%H
2025-12-04T08:53:14.0538760Z ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T08:53:14.0645109Z ##[group]Run actions/checkout@v4
2025-12-04T08:53:14.0645238Z with:
2025-12-04T08:53:14.0645361Z   ref: ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T08:53:14.0645489Z   fetch-depth: 0
2025-12-04T08:53:14.0645579Z   submodules: recursive
2025-12-04T08:53:14.0645686Z   show-progress: false
2025-12-04T08:53:14.0645820Z   repository: pytorch/pytorch
2025-12-04T08:53:14.0645970Z   token: ***
2025-12-04T08:53:14.0646057Z   ssh-strict: true
2025-12-04T08:53:14.0646143Z   ssh-user: git
2025-12-04T08:53:14.0646237Z   persist-credentials: true
2025-12-04T08:53:14.0646338Z   clean: true
2025-12-04T08:53:14.0646429Z   sparse-checkout-cone-mode: true
2025-12-04T08:53:14.0646543Z   fetch-tags: false
2025-12-04T08:53:14.0646627Z   lfs: false
2025-12-04T08:53:14.0646714Z   set-safe-directory: true
2025-12-04T08:53:14.0646811Z env:
2025-12-04T08:53:14.0646892Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:14.0646988Z ##[endgroup]
2025-12-04T08:53:14.1090800Z Syncing repository: pytorch/pytorch
2025-12-04T08:53:14.1091169Z ##[group]Getting Git version info
2025-12-04T08:53:14.1091330Z Working directory is '/home/runner/_work/pytorch/pytorch'
2025-12-04T08:53:14.1106650Z [command]/usr/bin/git version
2025-12-04T08:53:14.1131829Z git version 2.52.0
2025-12-04T08:53:14.1146318Z ##[endgroup]
2025-12-04T08:53:14.1151501Z Copying '/home/runner/.gitconfig' to '/home/runner/_work/_temp/f2a31106-c1b2-4781-b204-29006c329599/.gitconfig'
2025-12-04T08:53:14.1156577Z Temporarily overriding HOME='/home/runner/_work/_temp/f2a31106-c1b2-4781-b204-29006c329599' before making global git config changes
2025-12-04T08:53:14.1156898Z Adding repository directory to the temporary git global config as a safe directory
2025-12-04T08:53:14.1159368Z [command]/usr/bin/git config --global --add safe.directory /home/runner/_work/pytorch/pytorch
2025-12-04T08:53:14.1181864Z [command]/usr/bin/git config --local --get remote.origin.url
2025-12-04T08:53:14.1197078Z https://github.com/pytorch/pytorch
2025-12-04T08:53:14.1207164Z ##[group]Removing previously created refs, to avoid conflicts
2025-12-04T08:53:14.1209982Z [command]/usr/bin/git rev-parse --symbolic-full-name --verify --quiet HEAD
2025-12-04T08:53:14.1225356Z HEAD
2025-12-04T08:53:14.1251756Z ##[endgroup]
2025-12-04T08:53:14.1253323Z [command]/usr/bin/git submodule status
2025-12-04T08:53:14.1435002Z  7e1e1fe3858c63c251c637ae41a20de425dde96f android/libs/fbjni (v0.1.0-12-g7e1e1fe)
2025-12-04T08:53:14.1479394Z  4dfe081cf6bcd15db339cf2680b9281b8451eeb3 third_party/FP16 (4dfe081)
2025-12-04T08:53:14.1528915Z  b408327ac2a15ec3e43352421954f5b1967701d1 third_party/FXdiv (b408327)
2025-12-04T08:53:14.1587031Z  c07e3a0400713d546e0dea2d5466dd22ea389c73 third_party/NNPACK (c07e3a0)
2025-12-04T08:53:14.1636109Z  3ebbc93ded7285963bff932c678fa367eb393ba6 third_party/NVTX (v3.1.0-313-g3ebbc93)
2025-12-04T08:53:14.1696418Z  1d8f600fd424278486eade7ed3e877c99f0846b1 third_party/VulkanMemoryAllocator (v2.1.0-982-g1d8f600)
2025-12-04T08:53:14.1974146Z  51a0103656eff6fc9bfd39a4597923c4b542c883 third_party/XNNPACK (remotes/origin/ds/ndk-1243-g51a0103656)
2025-12-04T08:53:14.1997977Z  01aae101b9e5e94d6c16a9514c9fb8df99c93150 third_party/aiter (v0.1.1-92-g01aae101)
2025-12-04T08:53:14.2012013Z  299e5928955cc62af9968370293b916f5130916f third_party/benchmark (v1.9.3)
2025-12-04T08:53:14.2073224Z  7fe50dc3da2069d6645d9deb8c017a876472a977 third_party/composable_kernel (rocm-6.4.3-459-g7fe50dc3d)
2025-12-04T08:53:14.2148629Z  89c932f313c6437c38f2982869beacc89c2f2246 third_party/cpp-httplib (v0.26.0)
2025-12-04T08:53:14.2228211Z  f858c30bcb16f8effd5ff46996f0514539e17abc third_party/cpuinfo (f858c30)
2025-12-04T08:53:14.2263020Z  0b1577c8c83401237d601d0d0db5210506705396 third_party/cudnn_frontend (v0.5-61-g0b1577c)
2025-12-04T08:53:14.2345085Z  f88806b1e31dfa579842638740216dd41fc6c588 third_party/cutlass (v4.3.1)
2025-12-04T08:53:14.2363884Z  c0b988d39a9e47c794d699f29930ed4d7c7e13a4 third_party/fbgemm (v1.4.0-rc1-2-gc0b988d39)
2025-12-04T08:53:14.2422392Z  979702c87a8713a8e0a5e9fee122b90d2ef13be5 third_party/flash-attention (v2.7.4)
2025-12-04T08:53:14.2435237Z  a2cd1ea3b6d3fee220106b5fed3f7ce8da9eb757 third_party/flatbuffers (v24.12.23)
2025-12-04T08:53:14.2665125Z  407c905e45ad75fc29bf0f9bb7c5c2fd3475976f third_party/fmt (12.1.0)
2025-12-04T08:53:14.2721740Z  3fb5c176c17c765a3492cd2f0321b0dab712f350 third_party/gemmlowp/gemmlowp (remotes/origin/revert-87-master-135-g3fb5c17)
2025-12-04T08:53:14.2803420Z  54cbae0d3a67fa890b4c3d9ee162b7860315e341 third_party/gloo (remotes/origin/gh/c-p-i-o/1/base-37-g54cbae0)
2025-12-04T08:53:14.2939979Z  52eb8108c5bdec04579160ae17225d66034bd723 third_party/googletest (release-1.8.0-3544-g52eb8108)
2025-12-04T08:53:14.2996642Z  719d8e6cd7f7a0e01b155657526d693acf97c2b3 third_party/ideep (pytorch-rls-v3.7.1)
2025-12-04T08:53:14.3040028Z  dec1d23ca65ab069d225dfe40dea14f455170959 third_party/ittapi (v3.25.5)
2025-12-04T08:53:14.3162708Z  31f85df8fbd89c188f14ef10f1ec65379786b943 third_party/kineto (heads/main)
2025-12-04T08:53:14.3191430Z  d7770c89632329a9914ef1a90289917597639cbe third_party/kleidiai (v1.15.0)
2025-12-04T08:53:14.3210567Z  fbd8b99c2b828428947d70fdc046bb55609be93e third_party/mimalloc (v2.2.4)
2025-12-04T08:53:14.3224896Z  55f93686c01528224f448c19128836e7df245f72 third_party/nlohmann (v3.12.0)
2025-12-04T08:53:14.3441237Z  e709452ef2bbc1d113faf678c24e6d3467696e83 third_party/onnx (v1.18.0)
2025-12-04T08:53:14.3455161Z  a799f4aed9c94b765dcdaabaeab7d5e7e2310878 third_party/opentelemetry-cpp (v1.14.2)
2025-12-04T08:53:14.3483165Z  0fa0ef591e38c2758e3184c6c23e497b9f732ffa third_party/pocketfft (release_for_eigen-40-g0fa0ef5)
2025-12-04T08:53:14.3712714Z  d1eca4e4b421cd2997495c4b4e65cea6be4e9b8a third_party/protobuf (v3.7.0-rc.2-1279-gd1eca4e4b)
2025-12-04T08:53:14.3767293Z  072586a71b55b7f8c584153d223e95687148a900 third_party/psimd (heads/master)
2025-12-04T08:53:14.3813267Z  4fe0e1e183925bf8cfa6aae24237e724a96479b8 third_party/pthreadpool (0.1-144-g4fe0e1e)
2025-12-04T08:53:14.3832272Z  f5fbe867d2d26e4a0a9177a51f6e568868ad3dc8 third_party/pybind11 (v3.0.1)
2025-12-04T08:53:14.3907839Z  f45429b087dd7d5bc78bb40dc7cf06425c252d67 third_party/python-peachpy (remotes/origin/pre-generated)
2025-12-04T08:53:14.3975244Z  5a1d179df9cf652951b59010a2d2075372d67f68 third_party/sleef (3.8)
2025-12-04T08:53:14.4027728Z  2b4cd91092d335a697416b2a3cb398283246849d third_party/tensorpipe (heads/main)
2025-12-04T08:53:14.4037755Z ##[group]Cleaning the repository
2025-12-04T08:53:14.4041000Z [command]/usr/bin/git clean -ffdx
2025-12-04T08:53:14.4164903Z [command]/usr/bin/git reset --hard HEAD
2025-12-04T08:53:14.4877044Z HEAD is now at ffd9b0fb4355 Resolve collective autotuning test failure on arm (#168919)
2025-12-04T08:53:14.4935162Z ##[endgroup]
2025-12-04T08:53:14.4937782Z ##[group]Disabling automatic garbage collection
2025-12-04T08:53:14.4942921Z [command]/usr/bin/git config --local gc.auto 0
2025-12-04T08:53:14.4961751Z ##[endgroup]
2025-12-04T08:53:14.4961910Z ##[group]Setting up auth
2025-12-04T08:53:14.4965689Z [command]/usr/bin/git config --local --name-only --get-regexp core\.sshCommand
2025-12-04T08:53:14.4986913Z [command]/usr/bin/git submodule foreach --recursive sh -c "git config --local --name-only --get-regexp 'core\.sshCommand' && git config --local --unset-all 'core.sshCommand' || :"
2025-12-04T08:53:14.5184979Z Entering 'android/libs/fbjni'
2025-12-04T08:53:14.5208134Z Entering 'third_party/FP16'
2025-12-04T08:53:14.5228314Z Entering 'third_party/FXdiv'
2025-12-04T08:53:14.5248145Z Entering 'third_party/NNPACK'
2025-12-04T08:53:14.5271522Z Entering 'third_party/NVTX'
2025-12-04T08:53:14.5295254Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:14.5316221Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:14.5343318Z Entering 'third_party/aiter'
2025-12-04T08:53:14.5363254Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:14.5394678Z Entering 'third_party/benchmark'
2025-12-04T08:53:14.5422397Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:14.5447412Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:14.5471573Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:14.5493721Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:14.5520548Z Entering 'third_party/cutlass'
2025-12-04T08:53:14.5548839Z Entering 'third_party/fbgemm'
2025-12-04T08:53:14.5574764Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:14.5608198Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:14.5644505Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:14.5672824Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:14.5705676Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:14.5728035Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:14.5748099Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:14.5769982Z Entering 'third_party/flash-attention'
2025-12-04T08:53:14.5800024Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:14.5837331Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:14.5870935Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:14.5896380Z Entering 'third_party/fmt'
2025-12-04T08:53:14.5922536Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:14.5944697Z Entering 'third_party/gloo'
2025-12-04T08:53:14.5967584Z Entering 'third_party/googletest'
2025-12-04T08:53:14.5991624Z Entering 'third_party/ideep'
2025-12-04T08:53:14.6014769Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:14.6039049Z Entering 'third_party/ittapi'
2025-12-04T08:53:14.6062267Z Entering 'third_party/kineto'
2025-12-04T08:53:14.6086699Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:14.6108400Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:14.6128863Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:14.6151264Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:14.6174102Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:14.6195686Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:14.6219982Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:14.6243128Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:14.6264354Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:14.6284159Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:14.6311738Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:14.6335358Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:14.6366449Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:14.6394880Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:14.6417436Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:14.6442287Z Entering 'third_party/kleidiai'
2025-12-04T08:53:14.6468038Z Entering 'third_party/mimalloc'
2025-12-04T08:53:14.6491467Z Entering 'third_party/nlohmann'
2025-12-04T08:53:14.6515906Z Entering 'third_party/onnx'
2025-12-04T08:53:14.6544302Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:14.6570705Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:14.6599754Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:14.6630658Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:14.6653174Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:14.6674970Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:14.6696246Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:14.6718109Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:14.6741139Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:14.6765058Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:14.6789952Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:14.6814370Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:14.6846179Z Entering 'third_party/pocketfft'
2025-12-04T08:53:14.6872483Z Entering 'third_party/protobuf'
2025-12-04T08:53:14.6895085Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:14.6934048Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:14.6963503Z Entering 'third_party/psimd'
2025-12-04T08:53:14.6991507Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:14.7012181Z Entering 'third_party/pybind11'
2025-12-04T08:53:14.7037088Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:14.7064125Z Entering 'third_party/sleef'
2025-12-04T08:53:14.7086301Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:14.7109329Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:14.7134855Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:14.7157191Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:14.7178613Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:14.7197699Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:14.7236256Z [command]/usr/bin/git config --local --name-only --get-regexp http\.https\:\/\/github\.com\/\.extraheader
2025-12-04T08:53:14.7250808Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7257009Z [command]/usr/bin/git config --local --unset-all http.https://github.com/.extraheader
2025-12-04T08:53:14.7281162Z [command]/usr/bin/git submodule foreach --recursive sh -c "git config --local --name-only --get-regexp 'http\.https\:\/\/github\.com\/\.extraheader' && git config --local --unset-all 'http.https://github.com/.extraheader' || :"
2025-12-04T08:53:14.7438337Z Entering 'android/libs/fbjni'
2025-12-04T08:53:14.7457398Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7478231Z Entering 'third_party/FP16'
2025-12-04T08:53:14.7489702Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7506813Z Entering 'third_party/FXdiv'
2025-12-04T08:53:14.7519041Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7541985Z Entering 'third_party/NNPACK'
2025-12-04T08:53:14.7554531Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7579332Z Entering 'third_party/NVTX'
2025-12-04T08:53:14.7593528Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7612161Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:14.7625532Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7649961Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:14.7663055Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7685779Z Entering 'third_party/aiter'
2025-12-04T08:53:14.7700385Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7719243Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:14.7733332Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7757367Z Entering 'third_party/benchmark'
2025-12-04T08:53:14.7771355Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7790808Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:14.7803801Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7825637Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:14.7838844Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7854453Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:14.7866566Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7883759Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:14.7905237Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7928794Z Entering 'third_party/cutlass'
2025-12-04T08:53:14.7942831Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7964820Z Entering 'third_party/fbgemm'
2025-12-04T08:53:14.7977968Z http.https://github.com/.extraheader
2025-12-04T08:53:14.7992381Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:14.8009393Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8024673Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:14.8036387Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8055560Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:14.8070249Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8086452Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:14.8098774Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8118933Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:14.8132720Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8155659Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:14.8171384Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8189198Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:14.8206475Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8229151Z Entering 'third_party/flash-attention'
2025-12-04T08:53:14.8243012Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8260319Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:14.8274278Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8293260Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:14.8306558Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8339358Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:14.8361247Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8394797Z Entering 'third_party/fmt'
2025-12-04T08:53:14.8421025Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8448228Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:14.8479810Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8501704Z Entering 'third_party/gloo'
2025-12-04T08:53:14.8524539Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8542929Z Entering 'third_party/googletest'
2025-12-04T08:53:14.8562906Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8582696Z Entering 'third_party/ideep'
2025-12-04T08:53:14.8595908Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8618469Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:14.8632274Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8656553Z Entering 'third_party/ittapi'
2025-12-04T08:53:14.8668976Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8690010Z Entering 'third_party/kineto'
2025-12-04T08:53:14.8705699Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8725687Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:14.8738780Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8756137Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:14.8769438Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8786473Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:14.8802833Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8824354Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:14.8840601Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8857704Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:14.8870567Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8887327Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:14.8900802Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8925841Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:14.8937509Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8954130Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:14.8967282Z http.https://github.com/.extraheader
2025-12-04T08:53:14.8986395Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:14.8999753Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9017496Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:14.9030247Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9046360Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:14.9058986Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9077461Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:14.9090572Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9107658Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:14.9123675Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9146982Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:14.9160761Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9178219Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:14.9190840Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9210473Z Entering 'third_party/kleidiai'
2025-12-04T08:53:14.9224651Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9241825Z Entering 'third_party/mimalloc'
2025-12-04T08:53:14.9263677Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9279648Z Entering 'third_party/nlohmann'
2025-12-04T08:53:14.9296301Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9313619Z Entering 'third_party/onnx'
2025-12-04T08:53:14.9327207Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9351047Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:14.9369342Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9394089Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:14.9412171Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9430871Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:14.9443170Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9459630Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:14.9472965Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9487798Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:14.9498707Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9516021Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:14.9528804Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9544867Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:14.9557196Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9579947Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:14.9592613Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9609248Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:14.9620045Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9636622Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:14.9651520Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9670347Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:14.9682876Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9710077Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:14.9733562Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9762810Z Entering 'third_party/pocketfft'
2025-12-04T08:53:14.9776090Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9794189Z Entering 'third_party/protobuf'
2025-12-04T08:53:14.9808996Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9825725Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:14.9838675Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9863756Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:14.9877106Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9896986Z Entering 'third_party/psimd'
2025-12-04T08:53:14.9910737Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9929783Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:14.9944314Z http.https://github.com/.extraheader
2025-12-04T08:53:14.9967458Z Entering 'third_party/pybind11'
2025-12-04T08:53:14.9994911Z http.https://github.com/.extraheader
2025-12-04T08:53:15.0016136Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:15.0027817Z http.https://github.com/.extraheader
2025-12-04T08:53:15.0050929Z Entering 'third_party/sleef'
2025-12-04T08:53:15.0063957Z http.https://github.com/.extraheader
2025-12-04T08:53:15.0081235Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:15.0094490Z http.https://github.com/.extraheader
2025-12-04T08:53:15.0113440Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:15.0125682Z http.https://github.com/.extraheader
2025-12-04T08:53:15.0148619Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:15.0161300Z http.https://github.com/.extraheader
2025-12-04T08:53:15.0179218Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:15.0191539Z http.https://github.com/.extraheader
2025-12-04T08:53:15.0208830Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:15.0219907Z http.https://github.com/.extraheader
2025-12-04T08:53:15.0235482Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:15.0248816Z http.https://github.com/.extraheader
2025-12-04T08:53:15.0285168Z [command]/usr/bin/git config --local --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.0304731Z [command]/usr/bin/git submodule foreach --recursive git config --local --show-origin --name-only --get-regexp remote.origin.url
2025-12-04T08:53:15.0460458Z Entering 'android/libs/fbjni'
2025-12-04T08:53:15.0478677Z file:/home/runner/_work/pytorch/pytorch/.git/modules/android/libs/fbjni/config	remote.origin.url
2025-12-04T08:53:15.0489869Z Entering 'third_party/FP16'
2025-12-04T08:53:15.0500909Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FP16/config	remote.origin.url
2025-12-04T08:53:15.0509448Z Entering 'third_party/FXdiv'
2025-12-04T08:53:15.0519139Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FXdiv/config	remote.origin.url
2025-12-04T08:53:15.0527446Z Entering 'third_party/NNPACK'
2025-12-04T08:53:15.0537366Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK/config	remote.origin.url
2025-12-04T08:53:15.0545863Z Entering 'third_party/NVTX'
2025-12-04T08:53:15.0555778Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NVTX/config	remote.origin.url
2025-12-04T08:53:15.0564991Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:15.0575762Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/VulkanMemoryAllocator/config	remote.origin.url
2025-12-04T08:53:15.0584701Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:15.0594787Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/XNNPACK/config	remote.origin.url
2025-12-04T08:53:15.0609438Z Entering 'third_party/aiter'
2025-12-04T08:53:15.0620095Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/config	remote.origin.url
2025-12-04T08:53:15.0629289Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:15.0638548Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/modules/3rdparty/composable_kernel/config	remote.origin.url
2025-12-04T08:53:15.0652335Z Entering 'third_party/benchmark'
2025-12-04T08:53:15.0662045Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T08:53:15.0670843Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:15.0680589Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/composable_kernel/config	remote.origin.url
2025-12-04T08:53:15.0691842Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:15.0702463Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpp-httplib/config	remote.origin.url
2025-12-04T08:53:15.0711508Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:15.0721616Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpuinfo/config	remote.origin.url
2025-12-04T08:53:15.0730642Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:15.0742486Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cudnn_frontend/config	remote.origin.url
2025-12-04T08:53:15.0752541Z Entering 'third_party/cutlass'
2025-12-04T08:53:15.0762604Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cutlass/config	remote.origin.url
2025-12-04T08:53:15.0774721Z Entering 'third_party/fbgemm'
2025-12-04T08:53:15.0784107Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/config	remote.origin.url
2025-12-04T08:53:15.0794830Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:15.0809489Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/asmjit/config	remote.origin.url
2025-12-04T08:53:15.0818920Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:15.0828487Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/composable_kernel/config	remote.origin.url
2025-12-04T08:53:15.0840865Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:15.0849972Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cpuinfo/config	remote.origin.url
2025-12-04T08:53:15.0859129Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:15.0869424Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cutlass/config	remote.origin.url
2025-12-04T08:53:15.0881259Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:15.0890761Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/googletest/config	remote.origin.url
2025-12-04T08:53:15.0898988Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:15.0908667Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/hipify_torch/config	remote.origin.url
2025-12-04T08:53:15.0916334Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:15.0927271Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/json/config	remote.origin.url
2025-12-04T08:53:15.0938026Z Entering 'third_party/flash-attention'
2025-12-04T08:53:15.0948097Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/config	remote.origin.url
2025-12-04T08:53:15.0956613Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:15.0968607Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/composable_kernel/config	remote.origin.url
2025-12-04T08:53:15.0979508Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:15.0988563Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/cutlass/config	remote.origin.url
2025-12-04T08:53:15.1001177Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:15.1010844Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flatbuffers/config	remote.origin.url
2025-12-04T08:53:15.1020033Z Entering 'third_party/fmt'
2025-12-04T08:53:15.1029642Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fmt/config	remote.origin.url
2025-12-04T08:53:15.1038100Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:15.1047938Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/gemmlowp/gemmlowp/config	remote.origin.url
2025-12-04T08:53:15.1057195Z Entering 'third_party/gloo'
2025-12-04T08:53:15.1069824Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/gloo/config	remote.origin.url
2025-12-04T08:53:15.1079187Z Entering 'third_party/googletest'
2025-12-04T08:53:15.1088685Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:15.1097345Z Entering 'third_party/ideep'
2025-12-04T08:53:15.1107792Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/config	remote.origin.url
2025-12-04T08:53:15.1116021Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:15.1127107Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/modules/mkl-dnn/config	remote.origin.url
2025-12-04T08:53:15.1140360Z Entering 'third_party/ittapi'
2025-12-04T08:53:15.1150400Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ittapi/config	remote.origin.url
2025-12-04T08:53:15.1160879Z Entering 'third_party/kineto'
2025-12-04T08:53:15.1170489Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/config	remote.origin.url
2025-12-04T08:53:15.1180119Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:15.1194020Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/config	remote.origin.url
2025-12-04T08:53:15.1203578Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:15.1219614Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/DCGM/config	remote.origin.url
2025-12-04T08:53:15.1232363Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:15.1249547Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/cpr/config	remote.origin.url
2025-12-04T08:53:15.1264100Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:15.1275259Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/fmt/config	remote.origin.url
2025-12-04T08:53:15.1286225Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:15.1296187Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/config	remote.origin.url
2025-12-04T08:53:15.1305636Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:15.1314545Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/modules/doc/config	remote.origin.url
2025-12-04T08:53:15.1325582Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:15.1334879Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/glog/config	remote.origin.url
2025-12-04T08:53:15.1344057Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:15.1354213Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:15.1363217Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:15.1373169Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/json/config	remote.origin.url
2025-12-04T08:53:15.1384546Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:15.1394797Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/pfs/config	remote.origin.url
2025-12-04T08:53:15.1403998Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:15.1413631Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/config	remote.origin.url
2025-12-04T08:53:15.1423062Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:15.1435558Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/civetweb/config	remote.origin.url
2025-12-04T08:53:15.1445794Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:15.1458046Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/googletest/config	remote.origin.url
2025-12-04T08:53:15.1470489Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:15.1479663Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/fmt/config	remote.origin.url
2025-12-04T08:53:15.1488450Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:15.1499549Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:15.1515334Z Entering 'third_party/kleidiai'
2025-12-04T08:53:15.1525544Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kleidiai/config	remote.origin.url
2025-12-04T08:53:15.1535303Z Entering 'third_party/mimalloc'
2025-12-04T08:53:15.1545219Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/mimalloc/config	remote.origin.url
2025-12-04T08:53:15.1554916Z Entering 'third_party/nlohmann'
2025-12-04T08:53:15.1564525Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/nlohmann/config	remote.origin.url
2025-12-04T08:53:15.1574259Z Entering 'third_party/onnx'
2025-12-04T08:53:15.1584230Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/config	remote.origin.url
2025-12-04T08:53:15.1602603Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:15.1613180Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T08:53:15.1628020Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:15.1638464Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/config	remote.origin.url
2025-12-04T08:53:15.1647556Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:15.1656544Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T08:53:15.1664995Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:15.1674303Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:15.1683822Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:15.1693168Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/ms-gsl/config	remote.origin.url
2025-12-04T08:53:15.1703133Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:15.1712829Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/nlohmann-json/config	remote.origin.url
2025-12-04T08:53:15.1721574Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:15.1730994Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentelemetry-proto/config	remote.origin.url
2025-12-04T08:53:15.1744642Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:15.1754129Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentracing-cpp/config	remote.origin.url
2025-12-04T08:53:15.1762795Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:15.1774959Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/config	remote.origin.url
2025-12-04T08:53:15.1784436Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:15.1793535Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/civetweb/config	remote.origin.url
2025-12-04T08:53:15.1803049Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:15.1813455Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/googletest/config	remote.origin.url
2025-12-04T08:53:15.1829646Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:15.1839559Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/tools/vcpkg/config	remote.origin.url
2025-12-04T08:53:15.1857599Z Entering 'third_party/pocketfft'
2025-12-04T08:53:15.1868010Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/pocketfft/config	remote.origin.url
2025-12-04T08:53:15.1876571Z Entering 'third_party/protobuf'
2025-12-04T08:53:15.1887181Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/config	remote.origin.url
2025-12-04T08:53:15.1898814Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:15.1908944Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T08:53:15.1916958Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:15.1926043Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:15.1941499Z Entering 'third_party/psimd'
2025-12-04T08:53:15.1951644Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/psimd/config	remote.origin.url
2025-12-04T08:53:15.1961533Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:15.1973204Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/pthreadpool/config	remote.origin.url
2025-12-04T08:53:15.1982822Z Entering 'third_party/pybind11'
2025-12-04T08:53:15.1994143Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T08:53:15.2006101Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:15.2018653Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/python-peachpy/config	remote.origin.url
2025-12-04T08:53:15.2027957Z Entering 'third_party/sleef'
2025-12-04T08:53:15.2037788Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/sleef/config	remote.origin.url
2025-12-04T08:53:15.2046688Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:15.2056343Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/config	remote.origin.url
2025-12-04T08:53:15.2064937Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:15.2073715Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:15.2084687Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:15.2094498Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libnop/config	remote.origin.url
2025-12-04T08:53:15.2103980Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:15.2116970Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libuv/config	remote.origin.url
2025-12-04T08:53:15.2125596Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:15.2134912Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T08:53:15.2146450Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:15.2155545Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/modules/tools/clang/config	remote.origin.url
2025-12-04T08:53:15.2179469Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/android/libs/fbjni/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2198063Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FP16/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2212779Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FXdiv/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2228386Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2243012Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NVTX/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2263451Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/VulkanMemoryAllocator/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2277821Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/XNNPACK/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2293032Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2308492Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/modules/3rdparty/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2321824Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/benchmark/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2337745Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2352567Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpp-httplib/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2365201Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpuinfo/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2386801Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cudnn_frontend/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2403397Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cutlass/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2417223Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2431526Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/asmjit/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2445961Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2465108Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cpuinfo/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2485893Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cutlass/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2500017Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2514199Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/hipify_torch/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2528062Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/json/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2541838Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2557079Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2578699Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/cutlass/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2592669Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flatbuffers/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2606954Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fmt/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2621519Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/gemmlowp/gemmlowp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2636279Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/gloo/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2649568Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2664556Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2678544Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/modules/mkl-dnn/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2696720Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/ittapi/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2711183Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2723486Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2742390Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/DCGM/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2764246Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/cpr/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2779315Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/fmt/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2793616Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2808074Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/modules/doc/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2822337Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/glog/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2836251Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2850543Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/json/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2864906Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/pfs/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2878534Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2895014Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/civetweb/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2909706Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2923840Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/fmt/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2938212Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2952969Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kleidiai/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2966427Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/mimalloc/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.2987691Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/nlohmann/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3001929Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3016825Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/modules/third_party/pybind11/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3032842Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3048578Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/benchmark/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3062632Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3077069Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/ms-gsl/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3093762Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/nlohmann-json/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3108604Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentelemetry-proto/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3122564Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentracing-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3144644Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3159586Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/civetweb/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3174868Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3189587Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/tools/vcpkg/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3203971Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/pocketfft/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3218882Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3234719Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/benchmark/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3250498Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3265964Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/psimd/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3280612Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/pthreadpool/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3301328Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/pybind11/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3327876Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/python-peachpy/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3347724Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/sleef/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3363012Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3386851Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3406759Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libnop/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3423297Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libuv/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3439317Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3458418Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/modules/tools/clang/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T08:53:15.3485216Z [command]/usr/bin/git config --local http.https://github.com/.extraheader AUTHORIZATION: basic ***
2025-12-04T08:53:15.3508642Z ##[endgroup]
2025-12-04T08:53:15.3508837Z ##[group]Fetching the repository
2025-12-04T08:53:15.3512334Z [command]/usr/bin/git -c protocol.version=2 fetch --prune --no-recurse-submodules origin +refs/heads/*:refs/remotes/origin/* +refs/tags/*:refs/tags/*
2025-12-04T08:53:16.6077889Z [command]/usr/bin/git rev-parse --verify --quiet ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32^{object}
2025-12-04T08:53:16.6166600Z ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T08:53:16.6170735Z ##[endgroup]
2025-12-04T08:53:16.6171069Z ##[group]Determining the checkout info
2025-12-04T08:53:16.6172673Z ##[endgroup]
2025-12-04T08:53:16.6177470Z [command]/usr/bin/git sparse-checkout disable
2025-12-04T08:53:16.6272663Z [command]/usr/bin/git config --local --unset-all extensions.worktreeConfig
2025-12-04T08:53:16.6300761Z ##[group]Checking out the ref
2025-12-04T08:53:16.6302755Z [command]/usr/bin/git checkout --progress --force ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T08:53:16.6556246Z HEAD is now at ffd9b0fb4355 Resolve collective autotuning test failure on arm (#168919)
2025-12-04T08:53:16.6562396Z ##[endgroup]
2025-12-04T08:53:16.6562638Z ##[group]Setting up auth for fetching submodules
2025-12-04T08:53:16.6566271Z [command]/usr/bin/git config --global http.https://github.com/.extraheader AUTHORIZATION: basic ***
2025-12-04T08:53:16.6594695Z [command]/usr/bin/git config --global --unset-all url.https://github.com/.insteadOf
2025-12-04T08:53:16.6617846Z [command]/usr/bin/git config --global --add url.https://github.com/.insteadOf git@github.com:
2025-12-04T08:53:16.6634487Z [command]/usr/bin/git config --global --add url.https://github.com/.insteadOf org-21003710@github.com:
2025-12-04T08:53:16.6650199Z ##[endgroup]
2025-12-04T08:53:16.6650441Z ##[group]Fetching submodules
2025-12-04T08:53:16.6651890Z [command]/usr/bin/git submodule sync --recursive
2025-12-04T08:53:16.6850700Z Synchronizing submodule url for 'android/libs/fbjni'
2025-12-04T08:53:16.6864624Z Synchronizing submodule url for 'third_party/FP16'
2025-12-04T08:53:16.6877003Z Synchronizing submodule url for 'third_party/FXdiv'
2025-12-04T08:53:16.6888454Z Synchronizing submodule url for 'third_party/NNPACK'
2025-12-04T08:53:16.6899939Z Synchronizing submodule url for 'third_party/NVTX'
2025-12-04T08:53:16.6913553Z Synchronizing submodule url for 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:16.6924168Z Synchronizing submodule url for 'third_party/XNNPACK'
2025-12-04T08:53:16.6939675Z Synchronizing submodule url for 'third_party/aiter'
2025-12-04T08:53:16.6953712Z Synchronizing submodule url for 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:16.6967473Z Synchronizing submodule url for 'third_party/benchmark'
2025-12-04T08:53:16.6978536Z Synchronizing submodule url for 'third_party/composable_kernel'
2025-12-04T08:53:16.6998364Z Synchronizing submodule url for 'third_party/cpp-httplib'
2025-12-04T08:53:16.7010928Z Synchronizing submodule url for 'third_party/cpuinfo'
2025-12-04T08:53:16.7022151Z Synchronizing submodule url for 'third_party/cudnn_frontend'
2025-12-04T08:53:16.7040200Z Synchronizing submodule url for 'third_party/cutlass'
2025-12-04T08:53:16.7057950Z Synchronizing submodule url for 'third_party/fbgemm'
2025-12-04T08:53:16.7076196Z Synchronizing submodule url for 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:16.7085280Z Synchronizing submodule url for 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:16.7098125Z Synchronizing submodule url for 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:16.7107637Z Synchronizing submodule url for 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:16.7121344Z Synchronizing submodule url for 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:16.7130476Z Synchronizing submodule url for 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:16.7143688Z Synchronizing submodule url for 'third_party/fbgemm/external/json'
2025-12-04T08:53:16.7157150Z Synchronizing submodule url for 'third_party/flash-attention'
2025-12-04T08:53:16.7173404Z Synchronizing submodule url for 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:16.7186936Z Synchronizing submodule url for 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:16.7201511Z Synchronizing submodule url for 'third_party/flatbuffers'
2025-12-04T08:53:16.7215285Z Synchronizing submodule url for 'third_party/fmt'
2025-12-04T08:53:16.7232207Z Synchronizing submodule url for 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:16.7243430Z Synchronizing submodule url for 'third_party/gloo'
2025-12-04T08:53:16.7254517Z Synchronizing submodule url for 'third_party/googletest'
2025-12-04T08:53:16.7265307Z Synchronizing submodule url for 'third_party/ideep'
2025-12-04T08:53:16.7281727Z Synchronizing submodule url for 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:16.7297153Z Synchronizing submodule url for 'third_party/ittapi'
2025-12-04T08:53:16.7307914Z Synchronizing submodule url for 'third_party/kineto'
2025-12-04T08:53:16.7323613Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:16.7336160Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:16.7347762Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:16.7358776Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:16.7376287Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:16.7388829Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:16.7403353Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:16.7414637Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:16.7425803Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:16.7437130Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:16.7446425Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:16.7456321Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:16.7469923Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:16.7484710Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:16.7495252Z Synchronizing submodule url for 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:16.7515218Z Synchronizing submodule url for 'third_party/kleidiai'
2025-12-04T08:53:16.7526955Z Synchronizing submodule url for 'third_party/mimalloc'
2025-12-04T08:53:16.7537934Z Synchronizing submodule url for 'third_party/nlohmann'
2025-12-04T08:53:16.7549223Z Synchronizing submodule url for 'third_party/onnx'
2025-12-04T08:53:16.7577191Z Synchronizing submodule url for 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:16.7602349Z Synchronizing submodule url for 'third_party/opentelemetry-cpp'
2025-12-04T08:53:16.7615296Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:16.7625820Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:16.7637370Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:16.7649292Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:16.7660775Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:16.7678967Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:16.7694972Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:16.7705495Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:16.7726957Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:16.7740681Z Synchronizing submodule url for 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:16.7761114Z Synchronizing submodule url for 'third_party/pocketfft'
2025-12-04T08:53:16.7772323Z Synchronizing submodule url for 'third_party/protobuf'
2025-12-04T08:53:16.7785585Z Synchronizing submodule url for 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:16.7795977Z Synchronizing submodule url for 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:16.7810523Z Synchronizing submodule url for 'third_party/psimd'
2025-12-04T08:53:16.7821414Z Synchronizing submodule url for 'third_party/pthreadpool'
2025-12-04T08:53:16.7832195Z Synchronizing submodule url for 'third_party/pybind11'
2025-12-04T08:53:16.7843252Z Synchronizing submodule url for 'third_party/python-peachpy'
2025-12-04T08:53:16.7854690Z Synchronizing submodule url for 'third_party/sleef'
2025-12-04T08:53:16.7865149Z Synchronizing submodule url for 'third_party/tensorpipe'
2025-12-04T08:53:16.7877639Z Synchronizing submodule url for 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:16.7887743Z Synchronizing submodule url for 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:16.7900392Z Synchronizing submodule url for 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:16.7911726Z Synchronizing submodule url for 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:16.7922876Z Synchronizing submodule url for 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:16.7946622Z [command]/usr/bin/git -c protocol.version=2 submodule update --init --force --recursive
2025-12-04T08:53:16.8152504Z Submodule path 'android/libs/fbjni': checked out '7e1e1fe3858c63c251c637ae41a20de425dde96f'
2025-12-04T08:53:16.8195538Z Submodule path 'third_party/FP16': checked out '4dfe081cf6bcd15db339cf2680b9281b8451eeb3'
2025-12-04T08:53:16.8273641Z Submodule path 'third_party/FXdiv': checked out 'b408327ac2a15ec3e43352421954f5b1967701d1'
2025-12-04T08:53:16.8323149Z Submodule path 'third_party/NNPACK': checked out 'c07e3a0400713d546e0dea2d5466dd22ea389c73'
2025-12-04T08:53:16.8407440Z Submodule path 'third_party/NVTX': checked out '3ebbc93ded7285963bff932c678fa367eb393ba6'
2025-12-04T08:53:16.8475509Z Submodule path 'third_party/VulkanMemoryAllocator': checked out '1d8f600fd424278486eade7ed3e877c99f0846b1'
2025-12-04T08:53:16.8640816Z Submodule path 'third_party/XNNPACK': checked out '51a0103656eff6fc9bfd39a4597923c4b542c883'
2025-12-04T08:53:16.8772491Z Submodule path 'third_party/aiter': checked out '01aae101b9e5e94d6c16a9514c9fb8df99c93150'
2025-12-04T08:53:16.8934654Z Submodule path 'third_party/aiter/3rdparty/composable_kernel': checked out 'cffe8fa2a442ac8e80dd236a1a5d24fe3d7e0cbf'
2025-12-04T08:53:16.8996590Z Submodule path 'third_party/benchmark': checked out '299e5928955cc62af9968370293b916f5130916f'
2025-12-04T08:53:16.9179996Z Submodule path 'third_party/composable_kernel': checked out '7fe50dc3da2069d6645d9deb8c017a876472a977'
2025-12-04T08:53:16.9248133Z Submodule path 'third_party/cpp-httplib': checked out '89c932f313c6437c38f2982869beacc89c2f2246'
2025-12-04T08:53:16.9302348Z Submodule path 'third_party/cpuinfo': checked out 'f858c30bcb16f8effd5ff46996f0514539e17abc'
2025-12-04T08:53:16.9366140Z Submodule path 'third_party/cudnn_frontend': checked out '0b1577c8c83401237d601d0d0db5210506705396'
2025-12-04T08:53:16.9486565Z Submodule path 'third_party/cutlass': checked out 'f88806b1e31dfa579842638740216dd41fc6c588'
2025-12-04T08:53:16.9596951Z Submodule path 'third_party/fbgemm': checked out 'c0b988d39a9e47c794d699f29930ed4d7c7e13a4'
2025-12-04T08:53:16.9653873Z Submodule path 'third_party/fbgemm/external/asmjit': checked out 'a3199e8857792cd10b7589ff5d58343d2c9008ea'
2025-12-04T08:53:16.9817197Z Submodule path 'third_party/fbgemm/external/composable_kernel': checked out '7fe50dc3da2069d6645d9deb8c017a876472a977'
2025-12-04T08:53:16.9882635Z Submodule path 'third_party/fbgemm/external/cpuinfo': checked out '6543fec09b2f04ac4a666882998b534afc9c1349'
2025-12-04T08:53:16.9992001Z Submodule path 'third_party/fbgemm/external/cutlass': checked out '98125ce499b0fdf7ffbe0e3052f5b8709f4840f8'
2025-12-04T08:53:17.0052909Z Submodule path 'third_party/fbgemm/external/googletest': checked out '52eb8108c5bdec04579160ae17225d66034bd723'
2025-12-04T08:53:17.0108004Z Submodule path 'third_party/fbgemm/external/hipify_torch': checked out '63b6a7b541fa7f08f8475ca7d74054db36ff2691'
2025-12-04T08:53:17.0188291Z Submodule path 'third_party/fbgemm/external/json': checked out '9cca280a4d0ccf0c08f47a99aa71d1b0e52f8d03'
2025-12-04T08:53:17.0267699Z Submodule path 'third_party/flash-attention': checked out '979702c87a8713a8e0a5e9fee122b90d2ef13be5'
2025-12-04T08:53:17.0439640Z Submodule path 'third_party/flash-attention/csrc/composable_kernel': checked out '888317e698e9803c62bd38568abc9e05d7709f33'
2025-12-04T08:53:17.0559438Z Submodule path 'third_party/flash-attention/csrc/cutlass': checked out 'c506e16788cb08416a4a57e11a9067beeee29420'
2025-12-04T08:53:17.0650235Z Submodule path 'third_party/flatbuffers': checked out 'a2cd1ea3b6d3fee220106b5fed3f7ce8da9eb757'
2025-12-04T08:53:17.0708239Z Submodule path 'third_party/fmt': checked out '407c905e45ad75fc29bf0f9bb7c5c2fd3475976f'
2025-12-04T08:53:17.0758659Z Submodule path 'third_party/gemmlowp/gemmlowp': checked out '3fb5c176c17c765a3492cd2f0321b0dab712f350'
2025-12-04T08:53:17.0814917Z Submodule path 'third_party/gloo': checked out '54cbae0d3a67fa890b4c3d9ee162b7860315e341'
2025-12-04T08:53:17.0876228Z Submodule path 'third_party/googletest': checked out '52eb8108c5bdec04579160ae17225d66034bd723'
2025-12-04T08:53:17.0932125Z Submodule path 'third_party/ideep': checked out '719d8e6cd7f7a0e01b155657526d693acf97c2b3'
2025-12-04T08:53:17.1099905Z Submodule path 'third_party/ideep/mkl-dnn': checked out '8d263e693366ef8db40acc569cc7d8edf644556d'
2025-12-04T08:53:17.1153900Z Submodule path 'third_party/ittapi': checked out 'dec1d23ca65ab069d225dfe40dea14f455170959'
2025-12-04T08:53:17.1218947Z Submodule path 'third_party/kineto': checked out '31f85df8fbd89c188f14ef10f1ec65379786b943'
2025-12-04T08:53:17.1292700Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog': checked out 'd2ffe0a4e3acace628db49974246b66fc3e85fb1'
2025-12-04T08:53:17.1370931Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM': checked out 'ffde4e54bc7249a6039a5e6b45b395141e1217f9'
2025-12-04T08:53:17.1437808Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr': checked out '871ed52d350214a034f6ef8a3b8f51c5ce1bd400'
2025-12-04T08:53:17.1514886Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt': checked out 'cd4af11efc9c622896a3e4cb599fa28668ca3d05'
2025-12-04T08:53:17.1567268Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags': checked out 'e171aa2d15ed9eb17054558e0b3a6a413bb01067'
2025-12-04T08:53:17.1620090Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc': checked out '8411df715cf522606e3b1aca386ddfc0b63d34b4'
2025-12-04T08:53:17.1678234Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog': checked out 'b33e3bad4c46c8a6345525fd822af355e5ef9446'
2025-12-04T08:53:17.1736538Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest': checked out '52eb8108c5bdec04579160ae17225d66034bd723'
2025-12-04T08:53:17.1827127Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/json': checked out '4f8fba14066156b73f1189a2b8bd568bde5284c5'
2025-12-04T08:53:17.1880580Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs': checked out 'f68a2fa8ea36c783bdd760371411fcb495aa3150'
2025-12-04T08:53:17.1943136Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp': checked out 'b1234816facfdda29845c46696a02998a4af115a'
2025-12-04T08:53:17.2016135Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb': checked out 'd7ba35bbb649209c66e582d5a0244ba988a15159'
2025-12-04T08:53:17.2082046Z Submodule path 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest': checked out 'e2239ee6043f73722e7aa812a459f54a28552929'
2025-12-04T08:53:17.2137559Z Submodule path 'third_party/kineto/libkineto/third_party/fmt': checked out '40626af88bd7df9a5fb80be7b25ac85b122d6c21'
2025-12-04T08:53:17.2187036Z Submodule path 'third_party/kineto/libkineto/third_party/googletest': checked out '52eb8108c5bdec04579160ae17225d66034bd723'
2025-12-04T08:53:17.2257351Z Submodule path 'third_party/kleidiai': checked out 'd7770c89632329a9914ef1a90289917597639cbe'
2025-12-04T08:53:17.2329331Z Submodule path 'third_party/mimalloc': checked out 'fbd8b99c2b828428947d70fdc046bb55609be93e'
2025-12-04T08:53:17.2410185Z Submodule path 'third_party/nlohmann': checked out '55f93686c01528224f448c19128836e7df245f72'
2025-12-04T08:53:17.2573515Z Submodule path 'third_party/onnx': checked out 'e709452ef2bbc1d113faf678c24e6d3467696e83'
2025-12-04T08:53:17.2646863Z Submodule path 'third_party/onnx/third_party/pybind11': checked out 'a2e59f0e7065404b44dfe92a28aca47ba1378dc4'
2025-12-04T08:53:17.2742083Z Submodule path 'third_party/opentelemetry-cpp': checked out 'a799f4aed9c94b765dcdaabaeab7d5e7e2310878'
2025-12-04T08:53:17.2799383Z Submodule path 'third_party/opentelemetry-cpp/third_party/benchmark': checked out 'd572f4777349d43653b21d6c2fc63020ab326db2'
2025-12-04T08:53:17.2860730Z Submodule path 'third_party/opentelemetry-cpp/third_party/googletest': checked out 'b796f7d44681514f58a683a3a71ff17c94edb0c1'
2025-12-04T08:53:17.2909092Z Submodule path 'third_party/opentelemetry-cpp/third_party/ms-gsl': checked out '6f4529395c5b7c2d661812257cd6780c67e54afa'
2025-12-04T08:53:17.3002312Z Submodule path 'third_party/opentelemetry-cpp/third_party/nlohmann-json': checked out 'bc889afb4c5bf1c0d8ee29ef35eaaf4c8bef8a5d'
2025-12-04T08:53:17.3057077Z Submodule path 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto': checked out '4ca4f0335c63cda7ab31ea7ed70d6553aee14dce'
2025-12-04T08:53:17.3107899Z Submodule path 'third_party/opentelemetry-cpp/third_party/opentracing-cpp': checked out '06b57f48ded1fa3bdd3d4346f6ef29e40e08eaf5'
2025-12-04T08:53:17.3165331Z Submodule path 'third_party/opentelemetry-cpp/third_party/prometheus-cpp': checked out 'c9ffcdda9086ffd9e1283ea7a0276d831f3c8a8d'
2025-12-04T08:53:17.3236696Z Submodule path 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb': checked out 'eefb26f82b233268fc98577d265352720d477ba4'
2025-12-04T08:53:17.3298346Z Submodule path 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest': checked out 'e2239ee6043f73722e7aa812a459f54a28552929'
2025-12-04T08:53:17.3462657Z Submodule path 'third_party/opentelemetry-cpp/tools/vcpkg': checked out '8eb57355a4ffb410a2e94c07b4dca2dffbee8e50'
2025-12-04T08:53:17.3554888Z Submodule path 'third_party/pocketfft': checked out '0fa0ef591e38c2758e3184c6c23e497b9f732ffa'
2025-12-04T08:53:17.3725479Z Submodule path 'third_party/protobuf': checked out 'd1eca4e4b421cd2997495c4b4e65cea6be4e9b8a'
2025-12-04T08:53:17.3800487Z Submodule path 'third_party/protobuf/third_party/benchmark': checked out '5b7683f49e1e9223cf9927b24f6fd3d6bd82e3f8'
2025-12-04T08:53:17.3874405Z Submodule path 'third_party/protobuf/third_party/googletest': checked out '5ec7f0c4a113e2f18ac2c6cc7df51ad6afc24081'
2025-12-04T08:53:17.3927559Z Submodule path 'third_party/psimd': checked out '072586a71b55b7f8c584153d223e95687148a900'
2025-12-04T08:53:17.3979149Z Submodule path 'third_party/pthreadpool': checked out '4fe0e1e183925bf8cfa6aae24237e724a96479b8'
2025-12-04T08:53:17.4046894Z Submodule path 'third_party/pybind11': checked out 'f5fbe867d2d26e4a0a9177a51f6e568868ad3dc8'
2025-12-04T08:53:17.4098776Z Submodule path 'third_party/python-peachpy': checked out 'f45429b087dd7d5bc78bb40dc7cf06425c252d67'
2025-12-04T08:53:17.4152773Z Submodule path 'third_party/sleef': checked out '5a1d179df9cf652951b59010a2d2075372d67f68'
2025-12-04T08:53:17.4218547Z Submodule path 'third_party/tensorpipe': checked out '2b4cd91092d335a697416b2a3cb398283246849d'
2025-12-04T08:53:17.4268501Z Submodule path 'third_party/tensorpipe/third_party/googletest': checked out 'aee0f9d9b5b87796ee8a0ab26b7587ec30e8858e'
2025-12-04T08:53:17.4316299Z Submodule path 'third_party/tensorpipe/third_party/libnop': checked out '910b55815be16109f04f4180e9adee14fb4ce281'
2025-12-04T08:53:17.4451814Z Submodule path 'third_party/tensorpipe/third_party/libuv': checked out '5152db2cbfeb5582e9c27c5ea1dba2cd9e10759b'
2025-12-04T08:53:17.4517776Z Submodule path 'third_party/tensorpipe/third_party/pybind11': checked out 'a23996fce38ff6ccfbcdc09f1e63f2c4be5ea2ef'
2025-12-04T08:53:17.4578092Z Submodule path 'third_party/tensorpipe/third_party/pybind11/tools/clang': checked out '6a00cbc4a9b8e68b71caf7f774b3f9c753ae84d5'
2025-12-04T08:53:17.4613117Z [command]/usr/bin/git submodule foreach --recursive git config --local gc.auto 0
2025-12-04T08:53:17.4771395Z Entering 'android/libs/fbjni'
2025-12-04T08:53:17.4795429Z Entering 'third_party/FP16'
2025-12-04T08:53:17.4814847Z Entering 'third_party/FXdiv'
2025-12-04T08:53:17.4837177Z Entering 'third_party/NNPACK'
2025-12-04T08:53:17.4856884Z Entering 'third_party/NVTX'
2025-12-04T08:53:17.4875858Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:17.4894376Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:17.4921503Z Entering 'third_party/aiter'
2025-12-04T08:53:17.4943171Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:17.4967018Z Entering 'third_party/benchmark'
2025-12-04T08:53:17.4986751Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:17.5009881Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:17.5029249Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:17.5048908Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:17.5069867Z Entering 'third_party/cutlass'
2025-12-04T08:53:17.5096544Z Entering 'third_party/fbgemm'
2025-12-04T08:53:17.5117136Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:17.5137026Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:17.5159241Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:17.5180297Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:17.5201735Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:17.5220930Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:17.5239287Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:17.5267849Z Entering 'third_party/flash-attention'
2025-12-04T08:53:17.5291007Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:17.5313979Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:17.5345118Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:17.5368367Z Entering 'third_party/fmt'
2025-12-04T08:53:17.5390900Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:17.5422826Z Entering 'third_party/gloo'
2025-12-04T08:53:17.5443286Z Entering 'third_party/googletest'
2025-12-04T08:53:17.5463194Z Entering 'third_party/ideep'
2025-12-04T08:53:17.5482641Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:17.5504520Z Entering 'third_party/ittapi'
2025-12-04T08:53:17.5524676Z Entering 'third_party/kineto'
2025-12-04T08:53:17.5553273Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:17.5576521Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:17.5605430Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:17.5625492Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:17.5648539Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:17.5670396Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:17.5695220Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:17.5729084Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:17.5752524Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:17.5772537Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:17.5793094Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:17.5813374Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:17.5836385Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:17.5860897Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:17.5879925Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:17.5921080Z Entering 'third_party/kleidiai'
2025-12-04T08:53:17.5941332Z Entering 'third_party/mimalloc'
2025-12-04T08:53:17.5968800Z Entering 'third_party/nlohmann'
2025-12-04T08:53:17.5990171Z Entering 'third_party/onnx'
2025-12-04T08:53:17.6016686Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:17.6038663Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:17.6065545Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:17.6085228Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:17.6105085Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:17.6124493Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:17.6145151Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:17.6164300Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:17.6188018Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:17.6208180Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:17.6234309Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:17.6255409Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:17.6286176Z Entering 'third_party/pocketfft'
2025-12-04T08:53:17.6306153Z Entering 'third_party/protobuf'
2025-12-04T08:53:17.6326089Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:17.6347182Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:17.6368463Z Entering 'third_party/psimd'
2025-12-04T08:53:17.6388025Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:17.6409093Z Entering 'third_party/pybind11'
2025-12-04T08:53:17.6428872Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:17.6449834Z Entering 'third_party/sleef'
2025-12-04T08:53:17.6467841Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:17.6486613Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:17.6506708Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:17.6526506Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:17.6545932Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:17.6564509Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:17.6599575Z ##[endgroup]
2025-12-04T08:53:17.6599780Z ##[group]Persisting credentials for submodules
2025-12-04T08:53:17.6606008Z [command]/usr/bin/git submodule foreach --recursive sh -c "git config --local --name-only --get-regexp 'url\.https\:\/\/github\.com\/\.insteadOf' && git config --local --unset-all 'url.https://github.com/.insteadOf' || :"
2025-12-04T08:53:17.6769484Z Entering 'android/libs/fbjni'
2025-12-04T08:53:17.6786593Z url.https://github.com/.insteadof
2025-12-04T08:53:17.6786764Z url.https://github.com/.insteadof
2025-12-04T08:53:17.6812985Z Entering 'third_party/FP16'
2025-12-04T08:53:17.6836557Z url.https://github.com/.insteadof
2025-12-04T08:53:17.6836704Z url.https://github.com/.insteadof
2025-12-04T08:53:17.6858739Z Entering 'third_party/FXdiv'
2025-12-04T08:53:17.6874449Z url.https://github.com/.insteadof
2025-12-04T08:53:17.6874571Z url.https://github.com/.insteadof
2025-12-04T08:53:17.6893463Z Entering 'third_party/NNPACK'
2025-12-04T08:53:17.6906757Z url.https://github.com/.insteadof
2025-12-04T08:53:17.6906888Z url.https://github.com/.insteadof
2025-12-04T08:53:17.6925816Z Entering 'third_party/NVTX'
2025-12-04T08:53:17.6947377Z url.https://github.com/.insteadof
2025-12-04T08:53:17.6947510Z url.https://github.com/.insteadof
2025-12-04T08:53:17.6964123Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:17.6976167Z url.https://github.com/.insteadof
2025-12-04T08:53:17.6976322Z url.https://github.com/.insteadof
2025-12-04T08:53:17.6998047Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:17.7016137Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7016276Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7040056Z Entering 'third_party/aiter'
2025-12-04T08:53:17.7052975Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7053172Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7070821Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:17.7083446Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7083741Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7110290Z Entering 'third_party/benchmark'
2025-12-04T08:53:17.7124608Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7124868Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7147964Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:17.7162446Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7162692Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7191822Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:17.7207687Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7207925Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7223837Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:17.7241270Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7241485Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7264136Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:17.7276786Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7277001Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7293833Z Entering 'third_party/cutlass'
2025-12-04T08:53:17.7306892Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7307096Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7330119Z Entering 'third_party/fbgemm'
2025-12-04T08:53:17.7344875Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7345070Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7363759Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:17.7378930Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7379119Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7401315Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:17.7415697Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7416024Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7435610Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:17.7454686Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7454956Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7471126Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:17.7482575Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7482796Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7502933Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:17.7516700Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7516901Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7539874Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:17.7552734Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7552909Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7572823Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:17.7586598Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7586780Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7605730Z Entering 'third_party/flash-attention'
2025-12-04T08:53:17.7618742Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7618918Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7634969Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:17.7648195Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7648363Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7667938Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:17.7680683Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7680839Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7701089Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:17.7714576Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7714888Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7731826Z Entering 'third_party/fmt'
2025-12-04T08:53:17.7743409Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7743642Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7760591Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:17.7777632Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7777848Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7799075Z Entering 'third_party/gloo'
2025-12-04T08:53:17.7812529Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7812742Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7832168Z Entering 'third_party/googletest'
2025-12-04T08:53:17.7848283Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7848466Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7865605Z Entering 'third_party/ideep'
2025-12-04T08:53:17.7880205Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7880383Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7897939Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:17.7912087Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7912259Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7937247Z Entering 'third_party/ittapi'
2025-12-04T08:53:17.7954933Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7955258Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7974664Z Entering 'third_party/kineto'
2025-12-04T08:53:17.7987868Z url.https://github.com/.insteadof
2025-12-04T08:53:17.7988038Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8006618Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:17.8019231Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8019373Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8038733Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:17.8051375Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8051518Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8068536Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:17.8079944Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8080089Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8108241Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:17.8120742Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8138244Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8138439Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:17.8151971Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8152111Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8167878Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:17.8180697Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8180841Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8200934Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:17.8213287Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8213422Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8236680Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:17.8248298Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8265906Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8266087Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:17.8278286Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8278646Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8294884Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:17.8305972Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8306259Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8327505Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:17.8340485Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8340749Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8358216Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:17.8371035Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8371278Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8392140Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:17.8406183Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8406409Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8428414Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:17.8441799Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8442009Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8461155Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:17.8474114Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8474325Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8497770Z Entering 'third_party/kleidiai'
2025-12-04T08:53:17.8512886Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8513090Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8530705Z Entering 'third_party/mimalloc'
2025-12-04T08:53:17.8544371Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8544566Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8561821Z Entering 'third_party/nlohmann'
2025-12-04T08:53:17.8574774Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8574961Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8593369Z Entering 'third_party/onnx'
2025-12-04T08:53:17.8609693Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8609875Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8630775Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:17.8644576Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8644891Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8673410Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:17.8687682Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8687940Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8711348Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:17.8725922Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8726155Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8744988Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:17.8758653Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8758879Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8774846Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:17.8789896Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8790110Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8806430Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:17.8819790Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8819992Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8836014Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:17.8847434Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8847650Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8868216Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:17.8880770Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8880949Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8897275Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:17.8908146Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8908302Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8923835Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:17.8945161Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8945301Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8963029Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:17.8977982Z url.https://github.com/.insteadof
2025-12-04T08:53:17.8978109Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9000453Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:17.9012023Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9042217Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9042353Z Entering 'third_party/pocketfft'
2025-12-04T08:53:17.9060803Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9060944Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9077129Z Entering 'third_party/protobuf'
2025-12-04T08:53:17.9089370Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9110197Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9110347Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:17.9123194Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9123328Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9141163Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:17.9155710Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9155848Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9174450Z Entering 'third_party/psimd'
2025-12-04T08:53:17.9187290Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9187597Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9203853Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:17.9218737Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9218961Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9237725Z Entering 'third_party/pybind11'
2025-12-04T08:53:17.9251204Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9251332Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9268284Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:17.9281551Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9281741Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9298619Z Entering 'third_party/sleef'
2025-12-04T08:53:17.9311528Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9311642Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9333459Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:17.9347321Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9347447Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9368017Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:17.9385187Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9385304Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9402278Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:17.9413969Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9414141Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9429758Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:17.9440636Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9440759Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9458981Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:17.9470033Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9470156Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9486307Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:17.9498910Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9499024Z url.https://github.com/.insteadof
2025-12-04T08:53:17.9531950Z [command]/usr/bin/git submodule foreach --recursive sh -c "git config --local 'http.https://github.com/.extraheader' 'AUTHORIZATION: basic ***' && git config --local --show-origin --name-only --get-regexp remote.origin.url"
2025-12-04T08:53:17.9706853Z Entering 'android/libs/fbjni'
2025-12-04T08:53:17.9730108Z file:/home/runner/_work/pytorch/pytorch/.git/modules/android/libs/fbjni/config	remote.origin.url
2025-12-04T08:53:17.9741667Z Entering 'third_party/FP16'
2025-12-04T08:53:17.9762652Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FP16/config	remote.origin.url
2025-12-04T08:53:17.9773153Z Entering 'third_party/FXdiv'
2025-12-04T08:53:17.9796273Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FXdiv/config	remote.origin.url
2025-12-04T08:53:17.9805822Z Entering 'third_party/NNPACK'
2025-12-04T08:53:17.9825787Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK/config	remote.origin.url
2025-12-04T08:53:17.9836993Z Entering 'third_party/NVTX'
2025-12-04T08:53:17.9859536Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NVTX/config	remote.origin.url
2025-12-04T08:53:17.9873606Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:17.9895709Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/VulkanMemoryAllocator/config	remote.origin.url
2025-12-04T08:53:17.9905734Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:17.9925652Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/XNNPACK/config	remote.origin.url
2025-12-04T08:53:17.9940700Z Entering 'third_party/aiter'
2025-12-04T08:53:17.9965110Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/config	remote.origin.url
2025-12-04T08:53:17.9975891Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:17.9996229Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/modules/3rdparty/composable_kernel/config	remote.origin.url
2025-12-04T08:53:18.0009965Z Entering 'third_party/benchmark'
2025-12-04T08:53:18.0031176Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T08:53:18.0040000Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:18.0061646Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/composable_kernel/config	remote.origin.url
2025-12-04T08:53:18.0074987Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:18.0094641Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpp-httplib/config	remote.origin.url
2025-12-04T08:53:18.0105946Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:18.0126398Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpuinfo/config	remote.origin.url
2025-12-04T08:53:18.0136155Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:18.0158076Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cudnn_frontend/config	remote.origin.url
2025-12-04T08:53:18.0168267Z Entering 'third_party/cutlass'
2025-12-04T08:53:18.0186815Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cutlass/config	remote.origin.url
2025-12-04T08:53:18.0200296Z Entering 'third_party/fbgemm'
2025-12-04T08:53:18.0219729Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/config	remote.origin.url
2025-12-04T08:53:18.0232042Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:18.0260209Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/asmjit/config	remote.origin.url
2025-12-04T08:53:18.0270561Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:18.0294392Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/composable_kernel/config	remote.origin.url
2025-12-04T08:53:18.0307573Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:18.0332168Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cpuinfo/config	remote.origin.url
2025-12-04T08:53:18.0347744Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:18.0366892Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cutlass/config	remote.origin.url
2025-12-04T08:53:18.0379739Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:18.0401765Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/googletest/config	remote.origin.url
2025-12-04T08:53:18.0410815Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:18.0429159Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/hipify_torch/config	remote.origin.url
2025-12-04T08:53:18.0439085Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:18.0465070Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/json/config	remote.origin.url
2025-12-04T08:53:18.0478951Z Entering 'third_party/flash-attention'
2025-12-04T08:53:18.0501020Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/config	remote.origin.url
2025-12-04T08:53:18.0510752Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:18.0533389Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/composable_kernel/config	remote.origin.url
2025-12-04T08:53:18.0547952Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:18.0578498Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/cutlass/config	remote.origin.url
2025-12-04T08:53:18.0595342Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:18.0614212Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flatbuffers/config	remote.origin.url
2025-12-04T08:53:18.0625986Z Entering 'third_party/fmt'
2025-12-04T08:53:18.0645089Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fmt/config	remote.origin.url
2025-12-04T08:53:18.0655070Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:18.0676935Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/gemmlowp/gemmlowp/config	remote.origin.url
2025-12-04T08:53:18.0687580Z Entering 'third_party/gloo'
2025-12-04T08:53:18.0710781Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/gloo/config	remote.origin.url
2025-12-04T08:53:18.0720563Z Entering 'third_party/googletest'
2025-12-04T08:53:18.0741630Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:18.0751663Z Entering 'third_party/ideep'
2025-12-04T08:53:18.0777621Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/config	remote.origin.url
2025-12-04T08:53:18.0791185Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:18.0839889Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/modules/mkl-dnn/config	remote.origin.url
2025-12-04T08:53:18.0858862Z Entering 'third_party/ittapi'
2025-12-04T08:53:18.0877699Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ittapi/config	remote.origin.url
2025-12-04T08:53:18.0892720Z Entering 'third_party/kineto'
2025-12-04T08:53:18.0913929Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/config	remote.origin.url
2025-12-04T08:53:18.0922580Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:18.0942610Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/config	remote.origin.url
2025-12-04T08:53:18.0952494Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:18.0971567Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/DCGM/config	remote.origin.url
2025-12-04T08:53:18.0981671Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:18.1001209Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/cpr/config	remote.origin.url
2025-12-04T08:53:18.1009549Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:18.1029603Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/fmt/config	remote.origin.url
2025-12-04T08:53:18.1039354Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:18.1060578Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/config	remote.origin.url
2025-12-04T08:53:18.1069854Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:18.1090713Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/modules/doc/config	remote.origin.url
2025-12-04T08:53:18.1106686Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:18.1126207Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/glog/config	remote.origin.url
2025-12-04T08:53:18.1134400Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:18.1152009Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:18.1164464Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:18.1191143Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/json/config	remote.origin.url
2025-12-04T08:53:18.1201402Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:18.1222035Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/pfs/config	remote.origin.url
2025-12-04T08:53:18.1232451Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:18.1251238Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/config	remote.origin.url
2025-12-04T08:53:18.1260310Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:18.1280149Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/civetweb/config	remote.origin.url
2025-12-04T08:53:18.1290389Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:18.1309454Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/googletest/config	remote.origin.url
2025-12-04T08:53:18.1322455Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:18.1343388Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/fmt/config	remote.origin.url
2025-12-04T08:53:18.1356873Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:18.1376252Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:18.1390449Z Entering 'third_party/kleidiai'
2025-12-04T08:53:18.1413145Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kleidiai/config	remote.origin.url
2025-12-04T08:53:18.1425424Z Entering 'third_party/mimalloc'
2025-12-04T08:53:18.1448325Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/mimalloc/config	remote.origin.url
2025-12-04T08:53:18.1457725Z Entering 'third_party/nlohmann'
2025-12-04T08:53:18.1476362Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/nlohmann/config	remote.origin.url
2025-12-04T08:53:18.1488625Z Entering 'third_party/onnx'
2025-12-04T08:53:18.1508497Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/config	remote.origin.url
2025-12-04T08:53:18.1525867Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:18.1551038Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T08:53:18.1563788Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:18.1585211Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/config	remote.origin.url
2025-12-04T08:53:18.1594682Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:18.1615486Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T08:53:18.1625996Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:18.1647662Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:18.1658157Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:18.1683176Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/ms-gsl/config	remote.origin.url
2025-12-04T08:53:18.1692365Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:18.1712060Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/nlohmann-json/config	remote.origin.url
2025-12-04T08:53:18.1726382Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:18.1745080Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentelemetry-proto/config	remote.origin.url
2025-12-04T08:53:18.1757337Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:18.1779121Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentracing-cpp/config	remote.origin.url
2025-12-04T08:53:18.1789376Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:18.1808226Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/config	remote.origin.url
2025-12-04T08:53:18.1817196Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:18.1838158Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/civetweb/config	remote.origin.url
2025-12-04T08:53:18.1849243Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:18.1871404Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/googletest/config	remote.origin.url
2025-12-04T08:53:18.1881975Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:18.1901494Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/tools/vcpkg/config	remote.origin.url
2025-12-04T08:53:18.1920754Z Entering 'third_party/pocketfft'
2025-12-04T08:53:18.1940537Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/pocketfft/config	remote.origin.url
2025-12-04T08:53:18.1950546Z Entering 'third_party/protobuf'
2025-12-04T08:53:18.1970997Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/config	remote.origin.url
2025-12-04T08:53:18.1982145Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:18.2002851Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T08:53:18.2015103Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:18.2035298Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:18.2047213Z Entering 'third_party/psimd'
2025-12-04T08:53:18.2069642Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/psimd/config	remote.origin.url
2025-12-04T08:53:18.2079891Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:18.2100157Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/pthreadpool/config	remote.origin.url
2025-12-04T08:53:18.2110234Z Entering 'third_party/pybind11'
2025-12-04T08:53:18.2128318Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T08:53:18.2138327Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:18.2161500Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/python-peachpy/config	remote.origin.url
2025-12-04T08:53:18.2171311Z Entering 'third_party/sleef'
2025-12-04T08:53:18.2191290Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/sleef/config	remote.origin.url
2025-12-04T08:53:18.2201097Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:18.2223447Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/config	remote.origin.url
2025-12-04T08:53:18.2233778Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:18.2251491Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/googletest/config	remote.origin.url
2025-12-04T08:53:18.2260344Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:18.2279724Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libnop/config	remote.origin.url
2025-12-04T08:53:18.2289548Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:18.2309505Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libuv/config	remote.origin.url
2025-12-04T08:53:18.2319037Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:18.2338194Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T08:53:18.2347424Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:18.2369055Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/modules/tools/clang/config	remote.origin.url
2025-12-04T08:53:18.2582972Z [command]/usr/bin/git submodule foreach --recursive git config --local --add 'url.https://github.com/.insteadOf' 'git@github.com:'
2025-12-04T08:53:18.2748158Z Entering 'android/libs/fbjni'
2025-12-04T08:53:18.2769697Z Entering 'third_party/FP16'
2025-12-04T08:53:18.2789514Z Entering 'third_party/FXdiv'
2025-12-04T08:53:18.2808585Z Entering 'third_party/NNPACK'
2025-12-04T08:53:18.2831713Z Entering 'third_party/NVTX'
2025-12-04T08:53:18.2857659Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:18.2879728Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:18.2905671Z Entering 'third_party/aiter'
2025-12-04T08:53:18.2925365Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:18.2947963Z Entering 'third_party/benchmark'
2025-12-04T08:53:18.2971749Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:18.2994300Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:18.3015678Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:18.3035124Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:18.3053928Z Entering 'third_party/cutlass'
2025-12-04T08:53:18.3077177Z Entering 'third_party/fbgemm'
2025-12-04T08:53:18.3101401Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:18.3121745Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:18.3144402Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:18.3162438Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:18.3186392Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:18.3210921Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:18.3228614Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:18.3248516Z Entering 'third_party/flash-attention'
2025-12-04T08:53:18.3273081Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:18.3297373Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:18.3327428Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:18.3360493Z Entering 'third_party/fmt'
2025-12-04T08:53:18.3385135Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:18.3410161Z Entering 'third_party/gloo'
2025-12-04T08:53:18.3434278Z Entering 'third_party/googletest'
2025-12-04T08:53:18.3458518Z Entering 'third_party/ideep'
2025-12-04T08:53:18.3485437Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:18.3509510Z Entering 'third_party/ittapi'
2025-12-04T08:53:18.3531999Z Entering 'third_party/kineto'
2025-12-04T08:53:18.3551373Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:18.3571406Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:18.3610537Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:18.3630742Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:18.3651094Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:18.3676139Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:18.3698908Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:18.3717342Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:18.3735428Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:18.3754574Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:18.3773651Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:18.3792086Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:18.3813425Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:18.3838429Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:18.3857297Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:18.3877878Z Entering 'third_party/kleidiai'
2025-12-04T08:53:18.3897705Z Entering 'third_party/mimalloc'
2025-12-04T08:53:18.3916062Z Entering 'third_party/nlohmann'
2025-12-04T08:53:18.3943519Z Entering 'third_party/onnx'
2025-12-04T08:53:18.3971081Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:18.3996077Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:18.4018762Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:18.4039867Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:18.4063197Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:18.4081093Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:18.4102050Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:18.4120611Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:18.4139253Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:18.4161327Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:18.4182932Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:18.4215259Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:18.4241639Z Entering 'third_party/pocketfft'
2025-12-04T08:53:18.4261726Z Entering 'third_party/protobuf'
2025-12-04T08:53:18.4284403Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:18.4307224Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:18.4332677Z Entering 'third_party/psimd'
2025-12-04T08:53:18.4353092Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:18.4376030Z Entering 'third_party/pybind11'
2025-12-04T08:53:18.4396571Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:18.4418758Z Entering 'third_party/sleef'
2025-12-04T08:53:18.4439174Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:18.4460619Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:18.4481327Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:18.4500734Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:18.4522164Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:18.4544466Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:18.4583435Z [command]/usr/bin/git submodule foreach --recursive git config --local --add 'url.https://github.com/.insteadOf' 'org-21003710@github.com:'
2025-12-04T08:53:18.4750104Z Entering 'android/libs/fbjni'
2025-12-04T08:53:18.4771190Z Entering 'third_party/FP16'
2025-12-04T08:53:18.4793781Z Entering 'third_party/FXdiv'
2025-12-04T08:53:18.4813646Z Entering 'third_party/NNPACK'
2025-12-04T08:53:18.4832495Z Entering 'third_party/NVTX'
2025-12-04T08:53:18.4853126Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T08:53:18.4878169Z Entering 'third_party/XNNPACK'
2025-12-04T08:53:18.4904194Z Entering 'third_party/aiter'
2025-12-04T08:53:18.4925378Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T08:53:18.4949444Z Entering 'third_party/benchmark'
2025-12-04T08:53:18.4969012Z Entering 'third_party/composable_kernel'
2025-12-04T08:53:18.4992209Z Entering 'third_party/cpp-httplib'
2025-12-04T08:53:18.5011208Z Entering 'third_party/cpuinfo'
2025-12-04T08:53:18.5029612Z Entering 'third_party/cudnn_frontend'
2025-12-04T08:53:18.5048534Z Entering 'third_party/cutlass'
2025-12-04T08:53:18.5077285Z Entering 'third_party/fbgemm'
2025-12-04T08:53:18.5096968Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T08:53:18.5114350Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T08:53:18.5135002Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T08:53:18.5157506Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T08:53:18.5180961Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T08:53:18.5200657Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T08:53:18.5225440Z Entering 'third_party/fbgemm/external/json'
2025-12-04T08:53:18.5249827Z Entering 'third_party/flash-attention'
2025-12-04T08:53:18.5271431Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T08:53:18.5294493Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T08:53:18.5318210Z Entering 'third_party/flatbuffers'
2025-12-04T08:53:18.5340014Z Entering 'third_party/fmt'
2025-12-04T08:53:18.5362323Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T08:53:18.5385001Z Entering 'third_party/gloo'
2025-12-04T08:53:18.5407326Z Entering 'third_party/googletest'
2025-12-04T08:53:18.5428614Z Entering 'third_party/ideep'
2025-12-04T08:53:18.5448434Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T08:53:18.5472002Z Entering 'third_party/ittapi'
2025-12-04T08:53:18.5492364Z Entering 'third_party/kineto'
2025-12-04T08:53:18.5511545Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T08:53:18.5530614Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T08:53:18.5553651Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T08:53:18.5575311Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T08:53:18.5601869Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T08:53:18.5622407Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T08:53:18.5644909Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T08:53:18.5664716Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T08:53:18.5689139Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T08:53:18.5707823Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T08:53:18.5726227Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T08:53:18.5744699Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:18.5766765Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:18.5788699Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T08:53:18.5807340Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T08:53:18.5830725Z Entering 'third_party/kleidiai'
2025-12-04T08:53:18.5850118Z Entering 'third_party/mimalloc'
2025-12-04T08:53:18.5870906Z Entering 'third_party/nlohmann'
2025-12-04T08:53:18.5895610Z Entering 'third_party/onnx'
2025-12-04T08:53:18.5934657Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T08:53:18.5963146Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T08:53:18.5989628Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T08:53:18.6010862Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T08:53:18.6034452Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T08:53:18.6057870Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T08:53:18.6077586Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T08:53:18.6097741Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T08:53:18.6117068Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T08:53:18.6135466Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T08:53:18.6157054Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T08:53:18.6178522Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T08:53:18.6217336Z Entering 'third_party/pocketfft'
2025-12-04T08:53:18.6238788Z Entering 'third_party/protobuf'
2025-12-04T08:53:18.6259689Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T08:53:18.6278744Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T08:53:18.6301172Z Entering 'third_party/psimd'
2025-12-04T08:53:18.6322853Z Entering 'third_party/pthreadpool'
2025-12-04T08:53:18.6343015Z Entering 'third_party/pybind11'
2025-12-04T08:53:18.6364409Z Entering 'third_party/python-peachpy'
2025-12-04T08:53:18.6385208Z Entering 'third_party/sleef'
2025-12-04T08:53:18.6405527Z Entering 'third_party/tensorpipe'
2025-12-04T08:53:18.6424801Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T08:53:18.6445258Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T08:53:18.6466219Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T08:53:18.6484750Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T08:53:18.6502180Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T08:53:18.6532399Z ##[endgroup]
2025-12-04T08:53:18.6674261Z [command]/usr/bin/git log -1 --format=%H
2025-12-04T08:53:18.6755384Z ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T08:53:18.6888977Z Prepare all required actions
2025-12-04T08:53:18.6889300Z Getting action download info
2025-12-04T08:53:18.9246999Z Download action repository 'aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076' (SHA:062b18b96a7aff071d4dc91bc00c4c1a7945b076)
2025-12-04T08:53:19.7221422Z ##[group]Run ./.github/actions/setup-rocm
2025-12-04T08:53:19.7221568Z env:
2025-12-04T08:53:19.7221662Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:19.7221769Z ##[endgroup]
2025-12-04T08:53:19.7235348Z ##[group]Run dpkg -l | grep -E "  rocm"
2025-12-04T08:53:19.7235493Z dpkg -l | grep -E "  rocm"
2025-12-04T08:53:19.7239905Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:19.7240048Z env:
2025-12-04T08:53:19.7240132Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:19.7240243Z ##[endgroup]
2025-12-04T08:53:19.7303715Z ii  rocm-cmake                   0.14.0.60401-83~22.04                   amd64        rocm-cmake built using CMake
2025-12-04T08:53:19.7303992Z ii  rocm-core                    6.4.1.60401-83~22.04                    amd64        ROCm Runtime software stack
2025-12-04T08:53:19.7304237Z ii  rocm-dbgapi                  0.77.2.60401-83~22.04                   amd64        Library to provide AMD GPU debugger API
2025-12-04T08:53:19.7304489Z ii  rocm-debug-agent             2.0.4.60401-83~22.04                    amd64        Radeon Open Compute Debug Agent (ROCdebug-agent)
2025-12-04T08:53:19.7304751Z ii  rocm-dev                     6.4.1.60401-83~22.04                    amd64        Radeon Open Compute (ROCm) Runtime software stack
2025-12-04T08:53:19.7304996Z ii  rocm-device-libs             1.0.0.60401-83~22.04                    amd64        Radeon Open Compute - device libraries
2025-12-04T08:53:19.7305208Z ii  rocm-gdb                     15.2.60401-83~22.04                     amd64        ROCgdb
2025-12-04T08:53:19.7305409Z ii  rocm-llvm                    19.0.0.25184.60401-83~22.04             amd64        ROCm core compiler
2025-12-04T08:53:19.7305622Z ii  rocm-opencl                  2.0.0.60401-83~22.04                    amd64        clr built using CMake
2025-12-04T08:53:19.7305834Z ii  rocm-opencl-dev              2.0.0.60401-83~22.04                    amd64        clr built using CMake
2025-12-04T08:53:19.7306054Z ii  rocm-smi-lib                 7.5.0.60401-83~22.04                    amd64        AMD System Management libraries
2025-12-04T08:53:19.7306443Z ii  rocm-utils                   6.4.1.60401-83~22.04                    amd64        Radeon Open Compute (ROCm) Runtime software stack
2025-12-04T08:53:19.7306887Z ii  rocminfo                     1.0.0.60401-83~22.04                    amd64        Radeon Open Compute (ROCm) Runtime rocminfo tool
2025-12-04T08:53:19.7327760Z ##[group]Run # ignore expansion of "docker ps -q" since it could be empty
2025-12-04T08:53:19.7328127Z # ignore expansion of "docker ps -q" since it could be empty
2025-12-04T08:53:19.7328349Z # shellcheck disable=SC2046
2025-12-04T08:53:19.7328550Z docker stop $(docker ps -q) || true
2025-12-04T08:53:19.7328728Z # Prune all stopped containers.
2025-12-04T08:53:19.7328903Z docker container prune -f
2025-12-04T08:53:19.7334027Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:19.7334180Z env:
2025-12-04T08:53:19.7334295Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:19.7334409Z ##[endgroup]
2025-12-04T08:53:19.7554945Z docker: 'docker stop' requires at least 1 argument
2025-12-04T08:53:19.7555056Z 
2025-12-04T08:53:19.7555127Z Usage:  docker stop [OPTIONS] CONTAINER [CONTAINER...]
2025-12-04T08:53:19.7555225Z 
2025-12-04T08:53:19.7659518Z See 'docker stop --help' for more information
2025-12-04T08:53:19.7659652Z Total reclaimed space: 0B
2025-12-04T08:53:19.7688812Z ##[group]Run cat /etc/os-release || true
2025-12-04T08:53:19.7689077Z cat /etc/os-release || true
2025-12-04T08:53:19.7689288Z cat /etc/apt/sources.list.d/rocm.list || true
2025-12-04T08:53:19.7689704Z cat /opt/rocm/.info/version || true
2025-12-04T08:53:19.7689886Z whoami
2025-12-04T08:53:19.7694893Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:19.7695063Z env:
2025-12-04T08:53:19.7695165Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:19.7695282Z ##[endgroup]
2025-12-04T08:53:19.7714983Z PRETTY_NAME="Ubuntu 22.04.5 LTS"
2025-12-04T08:53:19.7715122Z NAME="Ubuntu"
2025-12-04T08:53:19.7715230Z VERSION_ID="22.04"
2025-12-04T08:53:19.7715331Z VERSION="22.04.5 LTS (Jammy Jellyfish)"
2025-12-04T08:53:19.7715454Z VERSION_CODENAME=jammy
2025-12-04T08:53:19.7715551Z ID=ubuntu
2025-12-04T08:53:19.7715632Z ID_LIKE=debian
2025-12-04T08:53:19.7715751Z HOME_URL="https://www.ubuntu.com/"
2025-12-04T08:53:19.7715878Z SUPPORT_URL="https://help.ubuntu.com/"
2025-12-04T08:53:19.7716032Z BUG_REPORT_URL="https://bugs.launchpad.net/ubuntu/"
2025-12-04T08:53:19.7716246Z PRIVACY_POLICY_URL="https://www.ubuntu.com/legal/terms-and-policies/privacy-policy"
2025-12-04T08:53:19.7716435Z UBUNTU_CODENAME=jammy
2025-12-04T08:53:19.7720613Z deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/6.4.1 jammy main
2025-12-04T08:53:19.7727448Z 6.4.1-83
2025-12-04T08:53:19.7735267Z runner
2025-12-04T08:53:19.7748481Z ##[group]Run dpkg -l | grep -E "  amdgpu"
2025-12-04T08:53:19.7748656Z dpkg -l | grep -E "  amdgpu"
2025-12-04T08:53:19.7752720Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:19.7752860Z env:
2025-12-04T08:53:19.7752946Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:19.7753046Z ##[endgroup]
2025-12-04T08:53:19.7810543Z ii  amdgpu-core                  1:6.4.60401-2164967.22.04               all          Core meta package for unified amdgpu driver.
2025-12-04T08:53:19.7810802Z ii  amdgpu-install               6.4.60401-2164967.22.04                 all          AMDGPU driver repository and installer
2025-12-04T08:53:19.7831509Z ##[group]Run rocm-smi
2025-12-04T08:53:19.7831722Z rocm-smi
2025-12-04T08:53:19.7836362Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:19.7836532Z env:
2025-12-04T08:53:19.7836638Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:19.7836758Z ##[endgroup]
2025-12-04T08:53:19.8388924Z 
2025-12-04T08:53:19.8389315Z 
2025-12-04T08:53:19.8389565Z ============================================ ROCm System Management Interface ============================================
2025-12-04T08:53:19.8390039Z ====================================================== Concise Info ======================================================
2025-12-04T08:53:19.8390293Z Device  Node  IDs              Temp        Power     Partitions          SCLK  MCLK    Fan  Perf    PwrCap   VRAM%  GPU%  
2025-12-04T08:53:19.8391017Z               (DID,     GUID)  (Junction)  (Socket)  (Mem, Compute, ID)                                                   
2025-12-04T08:53:19.8391234Z ==========================================================================================================================
2025-12-04T08:53:19.8391809Z 0       5     0x74a5,   2987   29.0°C      118.0W    NPS1, SPX, 0        N/A   900Mhz  0%   manual  1000.0W  0%     0%    
2025-12-04T08:53:19.8392026Z ==========================================================================================================================
2025-12-04T08:53:19.8392205Z ================================================== End of ROCm SMI Log ===================================================
2025-12-04T08:53:19.8455880Z ##[group]Run rocminfo
2025-12-04T08:53:19.8456034Z rocminfo
2025-12-04T08:53:19.8460672Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:19.8460854Z env:
2025-12-04T08:53:19.8460966Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:19.8461089Z ##[endgroup]
2025-12-04T08:53:19.9004480Z ROCk module version 6.12.12 is loaded
2025-12-04T08:53:19.9004733Z =====================    
2025-12-04T08:53:19.9004859Z HSA System Attributes    
2025-12-04T08:53:19.9004972Z =====================    
2025-12-04T08:53:19.9005285Z Runtime Version:         1.15
2025-12-04T08:53:19.9005415Z Runtime Ext Version:     1.7
2025-12-04T08:53:19.9005532Z System Timestamp Freq.:  1000.000000MHz
2025-12-04T08:53:19.9005716Z Sig. Max Wait Duration:  18446744073709551615 (0xFFFFFFFFFFFFFFFF) (timestamp count)
2025-12-04T08:53:19.9005919Z Machine Model:           LARGE                              
2025-12-04T08:53:19.9006095Z System Endianness:       LITTLE                             
2025-12-04T08:53:19.9006237Z Mwaitx:                  DISABLED
2025-12-04T08:53:19.9006351Z XNACK enabled:           NO
2025-12-04T08:53:19.9006458Z DMAbuf Support:          YES
2025-12-04T08:53:19.9006564Z VMM Support:             YES
2025-12-04T08:53:19.9006632Z 
2025-12-04T08:53:19.9006669Z ==========               
2025-12-04T08:53:19.9006770Z HSA Agents               
2025-12-04T08:53:19.9006866Z ==========               
2025-12-04T08:53:19.9006959Z *******                  
2025-12-04T08:53:19.9007053Z Agent 1                  
2025-12-04T08:53:19.9007148Z *******                  
2025-12-04T08:53:19.9007269Z   Name:                    AMD EPYC 9575F 64-Core Processor   
2025-12-04T08:53:19.9007447Z   Uuid:                    CPU-XX                             
2025-12-04T08:53:19.9007602Z   Marketing Name:          AMD EPYC 9575F 64-Core Processor   
2025-12-04T08:53:19.9007770Z   Vendor Name:             CPU                                
2025-12-04T08:53:19.9007923Z   Feature:                 None specified                     
2025-12-04T08:53:19.9008070Z   Profile:                 FULL_PROFILE                       
2025-12-04T08:53:19.9008224Z   Float Round Mode:        NEAR                               
2025-12-04T08:53:19.9008378Z   Max Queue Number:        0(0x0)                             
2025-12-04T08:53:19.9008529Z   Queue Min Size:          0(0x0)                             
2025-12-04T08:53:19.9008695Z   Queue Max Size:          0(0x0)                             
2025-12-04T08:53:19.9008847Z   Queue Type:              MULTI                              
2025-12-04T08:53:19.9008992Z   Node:                    0                                  
2025-12-04T08:53:19.9009139Z   Device Type:             CPU                                
2025-12-04T08:53:19.9009275Z   Cache Info:              
2025-12-04T08:53:19.9009392Z     L1:                      49152(0xc000) KB                   
2025-12-04T08:53:19.9009652Z   Chip ID:                 0(0x0)                             
2025-12-04T08:53:19.9009801Z   ASIC Revision:           0(0x0)                             
2025-12-04T08:53:19.9009958Z   Cacheline Size:          64(0x40)                           
2025-12-04T08:53:19.9010112Z   Max Clock Freq. (MHz):   3300                               
2025-12-04T08:53:19.9010259Z   BDFID:                   0                                  
2025-12-04T08:53:19.9010470Z   Internal Node ID:        0                                  
2025-12-04T08:53:19.9021369Z   Compute Unit:            64                                 
2025-12-04T08:53:19.9021543Z   SIMDs per CU:            0                                  
2025-12-04T08:53:19.9021697Z   Shader Engines:          0                                  
2025-12-04T08:53:19.9021847Z   Shader Arrs. per Eng.:   0                                  
2025-12-04T08:53:19.9022004Z   WatchPts on Addr. Ranges:1                                  
2025-12-04T08:53:19.9022150Z   Memory Properties:       
2025-12-04T08:53:19.9022256Z   Features:                None
2025-12-04T08:53:19.9022363Z   Pool Info:               
2025-12-04T08:53:19.9022466Z     Pool 1                   
2025-12-04T08:53:19.9022593Z       Segment:                 GLOBAL; FLAGS: FINE GRAINED        
2025-12-04T08:53:19.9022746Z       Size:                    1584733168(0x5e751bf0) KB          
2025-12-04T08:53:19.9022894Z       Allocatable:             TRUE                               
2025-12-04T08:53:19.9023046Z       Alloc Granule:           4KB                                
2025-12-04T08:53:19.9023269Z       Alloc Recommended Granule:4KB                                
2025-12-04T08:53:19.9023424Z       Alloc Alignment:         4KB                                
2025-12-04T08:53:19.9023579Z       Accessible by all:       TRUE                               
2025-12-04T08:53:19.9023720Z     Pool 2                   
2025-12-04T08:53:19.9023850Z       Segment:                 GLOBAL; FLAGS: EXTENDED FINE GRAINED
2025-12-04T08:53:19.9024001Z       Size:                    1584733168(0x5e751bf0) KB          
2025-12-04T08:53:19.9024141Z       Allocatable:             TRUE                               
2025-12-04T08:53:19.9024292Z       Alloc Granule:           4KB                                
2025-12-04T08:53:19.9024450Z       Alloc Recommended Granule:4KB                                
2025-12-04T08:53:19.9024604Z       Alloc Alignment:         4KB                                
2025-12-04T08:53:19.9024762Z       Accessible by all:       TRUE                               
2025-12-04T08:53:19.9024906Z     Pool 3                   
2025-12-04T08:53:19.9025030Z       Segment:                 GLOBAL; FLAGS: KERNARG, FINE GRAINED
2025-12-04T08:53:19.9025174Z       Size:                    1584733168(0x5e751bf0) KB          
2025-12-04T08:53:19.9025318Z       Allocatable:             TRUE                               
2025-12-04T08:53:19.9025471Z       Alloc Granule:           4KB                                
2025-12-04T08:53:19.9025631Z       Alloc Recommended Granule:4KB                                
2025-12-04T08:53:19.9025788Z       Alloc Alignment:         4KB                                
2025-12-04T08:53:19.9025951Z       Accessible by all:       TRUE                               
2025-12-04T08:53:19.9026089Z     Pool 4                   
2025-12-04T08:53:19.9026218Z       Segment:                 GLOBAL; FLAGS: COARSE GRAINED      
2025-12-04T08:53:19.9026372Z       Size:                    1584733168(0x5e751bf0) KB          
2025-12-04T08:53:19.9026533Z       Allocatable:             TRUE                               
2025-12-04T08:53:19.9026687Z       Alloc Granule:           4KB                                
2025-12-04T08:53:19.9026855Z       Alloc Recommended Granule:4KB                                
2025-12-04T08:53:19.9027016Z       Alloc Alignment:         4KB                                
2025-12-04T08:53:19.9027182Z       Accessible by all:       TRUE                               
2025-12-04T08:53:19.9027368Z   ISA Info:                
2025-12-04T08:53:19.9027477Z *******                  
2025-12-04T08:53:19.9027587Z Agent 2                  
2025-12-04T08:53:19.9027694Z *******                  
2025-12-04T08:53:19.9027815Z   Name:                    AMD EPYC 9575F 64-Core Processor   
2025-12-04T08:53:19.9027971Z   Uuid:                    CPU-XX                             
2025-12-04T08:53:19.9028127Z   Marketing Name:          AMD EPYC 9575F 64-Core Processor   
2025-12-04T08:53:19.9028295Z   Vendor Name:             CPU                                
2025-12-04T08:53:19.9028457Z   Feature:                 None specified                     
2025-12-04T08:53:19.9028609Z   Profile:                 FULL_PROFILE                       
2025-12-04T08:53:19.9028769Z   Float Round Mode:        NEAR                               
2025-12-04T08:53:19.9028931Z   Max Queue Number:        0(0x0)                             
2025-12-04T08:53:19.9029084Z   Queue Min Size:          0(0x0)                             
2025-12-04T08:53:19.9029240Z   Queue Max Size:          0(0x0)                             
2025-12-04T08:53:19.9029387Z   Queue Type:              MULTI                              
2025-12-04T08:53:19.9029536Z   Node:                    1                                  
2025-12-04T08:53:19.9029685Z   Device Type:             CPU                                
2025-12-04T08:53:19.9029820Z   Cache Info:              
2025-12-04T08:53:19.9029946Z     L1:                      49152(0xc000) KB                   
2025-12-04T08:53:19.9030121Z   Chip ID:                 0(0x0)                             
2025-12-04T08:53:19.9030268Z   ASIC Revision:           0(0x0)                             
2025-12-04T08:53:19.9030468Z   Cacheline Size:          64(0x40)                           
2025-12-04T08:53:19.9030621Z   Max Clock Freq. (MHz):   3300                               
2025-12-04T08:53:19.9030772Z   BDFID:                   0                                  
2025-12-04T08:53:19.9030930Z   Internal Node ID:        1                                  
2025-12-04T08:53:19.9031081Z   Compute Unit:            64                                 
2025-12-04T08:53:19.9031236Z   SIMDs per CU:            0                                  
2025-12-04T08:53:19.9031393Z   Shader Engines:          0                                  
2025-12-04T08:53:19.9031549Z   Shader Arrs. per Eng.:   0                                  
2025-12-04T08:53:19.9031717Z   WatchPts on Addr. Ranges:1                                  
2025-12-04T08:53:19.9031866Z   Memory Properties:       
2025-12-04T08:53:19.9031982Z   Features:                None
2025-12-04T08:53:19.9032101Z   Pool Info:               
2025-12-04T08:53:19.9032206Z     Pool 1                   
2025-12-04T08:53:19.9032344Z       Segment:                 GLOBAL; FLAGS: FINE GRAINED        
2025-12-04T08:53:19.9032503Z       Size:                    1585355648(0x5e7e9b80) KB          
2025-12-04T08:53:19.9032654Z       Allocatable:             TRUE                               
2025-12-04T08:53:19.9032818Z       Alloc Granule:           4KB                                
2025-12-04T08:53:19.9032988Z       Alloc Recommended Granule:4KB                                
2025-12-04T08:53:19.9033151Z       Alloc Alignment:         4KB                                
2025-12-04T08:53:19.9033316Z       Accessible by all:       TRUE                               
2025-12-04T08:53:19.9033453Z     Pool 2                   
2025-12-04T08:53:19.9033591Z       Segment:                 GLOBAL; FLAGS: EXTENDED FINE GRAINED
2025-12-04T08:53:19.9033754Z       Size:                    1585355648(0x5e7e9b80) KB          
2025-12-04T08:53:19.9033899Z       Allocatable:             TRUE                               
2025-12-04T08:53:19.9034058Z       Alloc Granule:           4KB                                
2025-12-04T08:53:19.9034228Z       Alloc Recommended Granule:4KB                                
2025-12-04T08:53:19.9034427Z       Alloc Alignment:         4KB                                
2025-12-04T08:53:19.9034593Z       Accessible by all:       TRUE                               
2025-12-04T08:53:19.9034730Z     Pool 3                   
2025-12-04T08:53:19.9034868Z       Segment:                 GLOBAL; FLAGS: KERNARG, FINE GRAINED
2025-12-04T08:53:19.9035023Z       Size:                    1585355648(0x5e7e9b80) KB          
2025-12-04T08:53:19.9035169Z       Allocatable:             TRUE                               
2025-12-04T08:53:19.9035328Z       Alloc Granule:           4KB                                
2025-12-04T08:53:19.9035488Z       Alloc Recommended Granule:4KB                                
2025-12-04T08:53:19.9035638Z       Alloc Alignment:         4KB                                
2025-12-04T08:53:19.9035786Z       Accessible by all:       TRUE                               
2025-12-04T08:53:19.9035914Z     Pool 4                   
2025-12-04T08:53:19.9036034Z       Segment:                 GLOBAL; FLAGS: COARSE GRAINED      
2025-12-04T08:53:19.9036178Z       Size:                    1585355648(0x5e7e9b80) KB          
2025-12-04T08:53:19.9036316Z       Allocatable:             TRUE                               
2025-12-04T08:53:19.9036460Z       Alloc Granule:           4KB                                
2025-12-04T08:53:19.9036614Z       Alloc Recommended Granule:4KB                                
2025-12-04T08:53:19.9036764Z       Alloc Alignment:         4KB                                
2025-12-04T08:53:19.9036914Z       Accessible by all:       TRUE                               
2025-12-04T08:53:19.9037043Z   ISA Info:                
2025-12-04T08:53:19.9037192Z *******                  
2025-12-04T08:53:19.9037285Z Agent 3                  
2025-12-04T08:53:19.9037376Z *******                  
2025-12-04T08:53:19.9037482Z   Name:                    gfx942                             
2025-12-04T08:53:19.9037619Z   Uuid:                    GPU-4158e280e9a05390               
2025-12-04T08:53:19.9037765Z   Marketing Name:          AMD Instinct MI325X                
2025-12-04T08:53:19.9037912Z   Vendor Name:             AMD                                
2025-12-04T08:53:19.9038055Z   Feature:                 KERNEL_DISPATCH                    
2025-12-04T08:53:19.9038199Z   Profile:                 BASE_PROFILE                       
2025-12-04T08:53:19.9038345Z   Float Round Mode:        NEAR                               
2025-12-04T08:53:19.9038489Z   Max Queue Number:        128(0x80)                          
2025-12-04T08:53:19.9038629Z   Queue Min Size:          64(0x40)                           
2025-12-04T08:53:19.9038777Z   Queue Max Size:          131072(0x20000)                    
2025-12-04T08:53:19.9038917Z   Queue Type:              MULTI                              
2025-12-04T08:53:19.9039049Z   Node:                    2                                  
2025-12-04T08:53:19.9039183Z   Device Type:             GPU                                
2025-12-04T08:53:19.9039312Z   Cache Info:              
2025-12-04T08:53:19.9039420Z     L1:                      32(0x20) KB                        
2025-12-04T08:53:19.9039544Z     L2:                      4096(0x1000) KB                    
2025-12-04T08:53:19.9039668Z     L3:                      262144(0x40000) KB                 
2025-12-04T08:53:19.9039797Z   Chip ID:                 29861(0x74a5)                      
2025-12-04T08:53:19.9039934Z   ASIC Revision:           1(0x1)                             
2025-12-04T08:53:19.9040077Z   Cacheline Size:          128(0x80)                          
2025-12-04T08:53:19.9040227Z   Max Clock Freq. (MHz):   2100                               
2025-12-04T08:53:19.9040362Z   BDFID:                   5376                               
2025-12-04T08:53:19.9040536Z   Internal Node ID:        2                                  
2025-12-04T08:53:19.9040676Z   Compute Unit:            304                                
2025-12-04T08:53:19.9040814Z   SIMDs per CU:            4                                  
2025-12-04T08:53:19.9040993Z   Shader Engines:          32                                 
2025-12-04T08:53:19.9041136Z   Shader Arrs. per Eng.:   1                                  
2025-12-04T08:53:19.9041285Z   WatchPts on Addr. Ranges:4                                  
2025-12-04T08:53:19.9041438Z   Coherent Host Access:    FALSE                              
2025-12-04T08:53:19.9041569Z   Memory Properties:       
2025-12-04T08:53:19.9041677Z   Features:                KERNEL_DISPATCH 
2025-12-04T08:53:19.9041809Z   Fast F16 Operation:      TRUE                               
2025-12-04T08:53:19.9041957Z   Wavefront Size:          64(0x40)                           
2025-12-04T08:53:19.9042106Z   Workgroup Max Size:      1024(0x400)                        
2025-12-04T08:53:19.9042240Z   Workgroup Max Size per Dimension:
2025-12-04T08:53:19.9042363Z     x                        1024(0x400)                        
2025-12-04T08:53:19.9042486Z     y                        1024(0x400)                        
2025-12-04T08:53:19.9042609Z     z                        1024(0x400)                        
2025-12-04T08:53:19.9042741Z   Max Waves Per CU:        32(0x20)                           
2025-12-04T08:53:19.9042894Z   Max Work-item Per CU:    2048(0x800)                        
2025-12-04T08:53:19.9043056Z   Grid Max Size:           4294967295(0xffffffff)             
2025-12-04T08:53:19.9043190Z   Grid Max Size per Dimension:
2025-12-04T08:53:19.9043298Z     x                        4294967295(0xffffffff)             
2025-12-04T08:53:19.9043424Z     y                        4294967295(0xffffffff)             
2025-12-04T08:53:19.9043585Z     z                        4294967295(0xffffffff)             
2025-12-04T08:53:19.9043725Z   Max fbarriers/Workgrp:   32                                 
2025-12-04T08:53:19.9048266Z   Packet Processor uCode:: 185                                
2025-12-04T08:53:19.9048435Z   SDMA engine uCode::      24                                 
2025-12-04T08:53:19.9048602Z   IOMMU Support::          None                               
2025-12-04T08:53:19.9048741Z   Pool Info:               
2025-12-04T08:53:19.9048842Z     Pool 1                   
2025-12-04T08:53:19.9048969Z       Segment:                 GLOBAL; FLAGS: COARSE GRAINED      
2025-12-04T08:53:19.9049117Z       Size:                    268419072(0xfffc000) KB            
2025-12-04T08:53:19.9049261Z       Allocatable:             TRUE                               
2025-12-04T08:53:19.9049412Z       Alloc Granule:           4KB                                
2025-12-04T08:53:19.9049574Z       Alloc Recommended Granule:2048KB                             
2025-12-04T08:53:19.9049727Z       Alloc Alignment:         4KB                                
2025-12-04T08:53:19.9049883Z       Accessible by all:       FALSE                              
2025-12-04T08:53:19.9050014Z     Pool 2                   
2025-12-04T08:53:19.9050138Z       Segment:                 GLOBAL; FLAGS: EXTENDED FINE GRAINED
2025-12-04T08:53:19.9050285Z       Size:                    268419072(0xfffc000) KB            
2025-12-04T08:53:19.9050469Z       Allocatable:             TRUE                               
2025-12-04T08:53:19.9050616Z       Alloc Granule:           4KB                                
2025-12-04T08:53:19.9050771Z       Alloc Recommended Granule:2048KB                             
2025-12-04T08:53:19.9050922Z       Alloc Alignment:         4KB                                
2025-12-04T08:53:19.9051072Z       Accessible by all:       FALSE                              
2025-12-04T08:53:19.9051199Z     Pool 3                   
2025-12-04T08:53:19.9051323Z       Segment:                 GLOBAL; FLAGS: FINE GRAINED        
2025-12-04T08:53:19.9051464Z       Size:                    268419072(0xfffc000) KB            
2025-12-04T08:53:19.9051601Z       Allocatable:             TRUE                               
2025-12-04T08:53:19.9051747Z       Alloc Granule:           4KB                                
2025-12-04T08:53:19.9051981Z       Alloc Recommended Granule:2048KB                             
2025-12-04T08:53:19.9052132Z       Alloc Alignment:         4KB                                
2025-12-04T08:53:19.9052281Z       Accessible by all:       FALSE                              
2025-12-04T08:53:19.9052408Z     Pool 4                   
2025-12-04T08:53:19.9052527Z       Segment:                 GROUP                              
2025-12-04T08:53:19.9052661Z       Size:                    64(0x40) KB                        
2025-12-04T08:53:19.9052799Z       Allocatable:             FALSE                              
2025-12-04T08:53:19.9052949Z       Alloc Granule:           0KB                                
2025-12-04T08:53:19.9053108Z       Alloc Recommended Granule:0KB                                
2025-12-04T08:53:19.9053259Z       Alloc Alignment:         0KB                                
2025-12-04T08:53:19.9053410Z       Accessible by all:       FALSE                              
2025-12-04T08:53:19.9053551Z   ISA Info:                
2025-12-04T08:53:19.9053652Z     ISA 1                    
2025-12-04T08:53:19.9053788Z       Name:                    amdgcn-amd-amdhsa--gfx942:sramecc+:xnack-
2025-12-04T08:53:19.9053950Z       Machine Models:          HSA_MACHINE_MODEL_LARGE            
2025-12-04T08:53:19.9054105Z       Profiles:                HSA_PROFILE_BASE                   
2025-12-04T08:53:19.9054265Z       Default Rounding Mode:   NEAR                               
2025-12-04T08:53:19.9054428Z       Default Rounding Mode:   NEAR                               
2025-12-04T08:53:19.9054620Z       Fast f16:                TRUE                               
2025-12-04T08:53:19.9054779Z       Workgroup Max Size:      1024(0x400)                        
2025-12-04T08:53:19.9054923Z       Workgroup Max Size per Dimension:
2025-12-04T08:53:19.9055059Z         x                        1024(0x400)                        
2025-12-04T08:53:19.9055190Z         y                        1024(0x400)                        
2025-12-04T08:53:19.9055327Z         z                        1024(0x400)                        
2025-12-04T08:53:19.9055473Z       Grid Max Size:           4294967295(0xffffffff)             
2025-12-04T08:53:19.9055608Z       Grid Max Size per Dimension:
2025-12-04T08:53:19.9055735Z         x                        4294967295(0xffffffff)             
2025-12-04T08:53:19.9055870Z         y                        4294967295(0xffffffff)             
2025-12-04T08:53:19.9055999Z         z                        4294967295(0xffffffff)             
2025-12-04T08:53:19.9056147Z       FBarrier Max Size:       32                                 
2025-12-04T08:53:19.9056274Z     ISA 2                    
2025-12-04T08:53:19.9056410Z       Name:                    amdgcn-amd-amdhsa--gfx9-4-generic:sramecc+:xnack-
2025-12-04T08:53:19.9056577Z       Machine Models:          HSA_MACHINE_MODEL_LARGE            
2025-12-04T08:53:19.9056731Z       Profiles:                HSA_PROFILE_BASE                   
2025-12-04T08:53:19.9056888Z       Default Rounding Mode:   NEAR                               
2025-12-04T08:53:19.9057047Z       Default Rounding Mode:   NEAR                               
2025-12-04T08:53:19.9057191Z       Fast f16:                TRUE                               
2025-12-04T08:53:19.9057338Z       Workgroup Max Size:      1024(0x400)                        
2025-12-04T08:53:19.9057475Z       Workgroup Max Size per Dimension:
2025-12-04T08:53:19.9057594Z         x                        1024(0x400)                        
2025-12-04T08:53:19.9057716Z         y                        1024(0x400)                        
2025-12-04T08:53:19.9057838Z         z                        1024(0x400)                        
2025-12-04T08:53:19.9057975Z       Grid Max Size:           4294967295(0xffffffff)             
2025-12-04T08:53:19.9058106Z       Grid Max Size per Dimension:
2025-12-04T08:53:19.9058216Z         x                        4294967295(0xffffffff)             
2025-12-04T08:53:19.9058367Z         y                        4294967295(0xffffffff)             
2025-12-04T08:53:19.9058490Z         z                        4294967295(0xffffffff)             
2025-12-04T08:53:19.9058623Z       FBarrier Max Size:       32                                 
2025-12-04T08:53:19.9058752Z *** Done ***             
2025-12-04T08:53:19.9067452Z ##[group]Run ngpu=$(rocminfo | grep -c -E 'Name:.*\sgfx')
2025-12-04T08:53:19.9067636Z ngpu=$(rocminfo | grep -c -E 'Name:.*\sgfx')
2025-12-04T08:53:19.9067913Z msg="Please file an issue on pytorch/pytorch reporting the faulty runner. Include a link to the runner logs so the runner can be identified"
2025-12-04T08:53:19.9068175Z if [[ $ngpu -eq 0 ]]; then
2025-12-04T08:53:19.9068318Z     echo "Error: Failed to detect any GPUs on the runner"
2025-12-04T08:53:19.9068455Z     echo "$msg"
2025-12-04T08:53:19.9068553Z     exit 1
2025-12-04T08:53:19.9068642Z fi
2025-12-04T08:53:19.9071407Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:19.9071550Z env:
2025-12-04T08:53:19.9071636Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:19.9071738Z ##[endgroup]
2025-12-04T08:53:19.9602605Z ##[group]Run pytorch/pytorch/.github/actions/diskspace-cleanup@main
2025-12-04T08:53:19.9602788Z with:
2025-12-04T08:53:19.9602888Z   diskspace-cutoff: 70
2025-12-04T08:53:19.9602991Z env:
2025-12-04T08:53:19.9603084Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:19.9603197Z ##[endgroup]
2025-12-04T08:53:19.9626152Z ##[group]Run set -ex
2025-12-04T08:53:19.9626311Z set -ex
2025-12-04T08:53:19.9626580Z diskspace_cutoff=70
2025-12-04T08:53:19.9626736Z docker_root_dir=$(docker info -f '{{.DockerRootDir}}')
2025-12-04T08:53:19.9626903Z if [ ! -d "$docker_root_dir" ]; then
2025-12-04T08:53:19.9627099Z     echo "Docker root directory ($docker_root_dir) does not exist. Skipping disk space check."
2025-12-04T08:53:19.9627304Z     exit 0
2025-12-04T08:53:19.9627405Z fi
2025-12-04T08:53:19.9627568Z diskspace=$(df -H --output=pcent ${docker_root_dir} | sed -n 2p | sed 's/%//' | sed 's/ //')
2025-12-04T08:53:19.9627895Z msg="Please file an issue on pytorch/pytorch reporting the faulty runner. Include a link to the runner logs so the runner can be identified"
2025-12-04T08:53:19.9628179Z if [[ "$diskspace" -ge "$diskspace_cutoff" ]] ; then
2025-12-04T08:53:19.9628325Z     docker system prune -af
2025-12-04T08:53:19.9628542Z     diskspace_new=$(df -H --output=pcent ${docker_root_dir} | sed -n 2p | sed 's/%//' | sed 's/ //')
2025-12-04T08:53:19.9628758Z     if [[ "$diskspace_new" -gt "$diskspace_cutoff" ]] ; then
2025-12-04T08:53:19.9628920Z         diskspace_cutoff_int=$((diskspace_cutoff + 0))
2025-12-04T08:53:19.9629081Z         difference=$((100 - diskspace_cutoff_int))
2025-12-04T08:53:19.9629287Z         echo "Error: Available diskspace is less than $difference percent. Not enough diskspace."
2025-12-04T08:53:19.9629477Z         echo "$msg"
2025-12-04T08:53:19.9629584Z         exit 1
2025-12-04T08:53:19.9629686Z     else
2025-12-04T08:53:19.9629796Z         difference=$((diskspace - diskspace_new))
2025-12-04T08:53:19.9629950Z         echo "Diskspace saved: $difference percent"
2025-12-04T08:53:19.9630084Z     fi
2025-12-04T08:53:19.9630168Z fi
2025-12-04T08:53:19.9634217Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:19.9634362Z env:
2025-12-04T08:53:19.9634461Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:19.9634565Z ##[endgroup]
2025-12-04T08:53:19.9646334Z + diskspace_cutoff=70
2025-12-04T08:53:19.9649449Z ++ docker info -f '{{.DockerRootDir}}'
2025-12-04T08:53:19.9959014Z + docker_root_dir=/home/runner/docker-data
2025-12-04T08:53:19.9959166Z + '[' '!' -d /home/runner/docker-data ']'
2025-12-04T08:53:19.9966742Z ++ df -H --output=pcent /home/runner/docker-data
2025-12-04T08:53:19.9967729Z ++ sed -n 2p
2025-12-04T08:53:19.9969122Z ++ sed s/%//
2025-12-04T08:53:19.9969914Z ++ sed 's/ //'
2025-12-04T08:53:19.9986547Z + diskspace=' 4'
2025-12-04T08:53:19.9987225Z + msg='Please file an issue on pytorch/pytorch reporting the faulty runner. Include a link to the runner logs so the runner can be identified'
2025-12-04T08:53:19.9987724Z + [[  4 -ge 70 ]]
2025-12-04T08:53:20.0018393Z ##[group]Run RUNNER_ARTIFACT_DIR="${RUNNER_TEMP}/artifacts"
2025-12-04T08:53:20.0018669Z RUNNER_ARTIFACT_DIR="${RUNNER_TEMP}/artifacts"
2025-12-04T08:53:20.0018879Z rm -rf "${RUNNER_ARTIFACT_DIR}"
2025-12-04T08:53:20.0019055Z mkdir -p "${RUNNER_ARTIFACT_DIR}"
2025-12-04T08:53:20.0019270Z echo "RUNNER_ARTIFACT_DIR=${RUNNER_ARTIFACT_DIR}" >> "${GITHUB_ENV}"
2025-12-04T08:53:20.0019469Z 
2025-12-04T08:53:20.0019620Z RUNNER_TEST_RESULTS_DIR="${RUNNER_TEMP}/test-results"
2025-12-04T08:53:20.0019829Z rm -rf "${RUNNER_TEST_RESULTS_DIR}"
2025-12-04T08:53:20.0020003Z mkdir -p "${RUNNER_TEST_RESULTS_DIR}"
2025-12-04T08:53:20.0020229Z echo "RUNNER_TEST_RESULTS_DIR=${RUNNER_TEST_RESULTS_DIR}" >> "${GITHUB_ENV}"
2025-12-04T08:53:20.0020490Z 
2025-12-04T08:53:20.0020617Z RUNNER_DOCS_DIR="${RUNNER_TEMP}/docs"
2025-12-04T08:53:20.0020779Z rm -rf "${RUNNER_DOCS_DIR}"
2025-12-04T08:53:20.0020931Z mkdir -p "${RUNNER_DOCS_DIR}"
2025-12-04T08:53:20.0021119Z echo "RUNNER_DOCS_DIR=${RUNNER_DOCS_DIR}" >> "${GITHUB_ENV}"
2025-12-04T08:53:20.0025572Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:20.0025715Z env:
2025-12-04T08:53:20.0025805Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:20.0025907Z ##[endgroup]
2025-12-04T08:53:20.0086437Z ##[group]Run env | grep '^GITHUB' >> "${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}"
2025-12-04T08:53:20.0086683Z env | grep '^GITHUB' >> "${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}"
2025-12-04T08:53:20.0086911Z env | grep '^CI' >> "${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}"
2025-12-04T08:53:20.0090349Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:20.0090616Z env:
2025-12-04T08:53:20.0090719Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:20.0090865Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T08:53:20.0091063Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T08:53:20.0091247Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T08:53:20.0091386Z ##[endgroup]
2025-12-04T08:53:20.0134145Z ##[group]Run # All GPUs are visible to the runner; visibility, if needed, will be set by run_test.py.
2025-12-04T08:53:20.0134454Z # All GPUs are visible to the runner; visibility, if needed, will be set by run_test.py.
2025-12-04T08:53:20.0134672Z # Add render group for container creation.
2025-12-04T08:53:20.0134854Z render_gid=`cat /etc/group | grep render | cut -d: -f3`
2025-12-04T08:53:20.0135082Z # Ensure GPU isolation if pod is part of kubernetes setup with DEVICE_FLAG.
2025-12-04T08:53:20.0135301Z if [ -f "/etc/podinfo/gha-render-devices" ]; then
2025-12-04T08:53:20.0135479Z   DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
2025-12-04T08:53:20.0135635Z else
2025-12-04T08:53:20.0135743Z   DEVICE_FLAG="--device /dev/dri"
2025-12-04T08:53:20.0135872Z fi
2025-12-04T08:53:20.0136067Z # The --group-add daemon and --group-add bin are needed in the Ubuntu 24.04 and Almalinux OSs respectively.
2025-12-04T08:53:20.0136357Z # This is due to the device files (/dev/kfd & /dev/dri) being owned by video group on bare metal.
2025-12-04T08:53:20.0136625Z # This video group ID maps to subgid 1 inside the docker image due to the /etc/subgid entries.
2025-12-04T08:53:20.0136905Z # The group name corresponding to group ID 1 can change depending on the OS, so both are necessary.
2025-12-04T08:53:20.0137498Z echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd $DEVICE_FLAG --group-add video --group-add $render_gid --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host" >> "${GITHUB_ENV}"
2025-12-04T08:53:20.0140859Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:20.0141014Z env:
2025-12-04T08:53:20.0141111Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:20.0141252Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T08:53:20.0141439Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T08:53:20.0141599Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T08:53:20.0141729Z ##[endgroup]
2025-12-04T08:53:20.0203315Z ##[group]Run aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722
2025-12-04T08:53:20.0203519Z with:
2025-12-04T08:53:20.0203671Z   role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
2025-12-04T08:53:20.0203863Z   aws-region: us-east-1
2025-12-04T08:53:20.0203975Z   role-duration-seconds: 18000
2025-12-04T08:53:20.0204101Z   audience: sts.amazonaws.com
2025-12-04T08:53:20.0204209Z env:
2025-12-04T08:53:20.0204299Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:20.0204432Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T08:53:20.0204607Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T08:53:20.0204773Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T08:53:20.0205281Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T08:53:20.0205641Z ##[endgroup]
2025-12-04T08:53:20.3270480Z Assuming role with OIDC
2025-12-04T08:53:20.6528029Z Authenticated as assumedRoleId AROAUPVRELQNLLCOPFEJR:GitHubActions
2025-12-04T08:53:20.7430136Z ##[group]Run aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076
2025-12-04T08:53:20.7430364Z with:
2025-12-04T08:53:20.7430536Z   mask-password: true
2025-12-04T08:53:20.7430671Z   registry-type: private
2025-12-04T08:53:20.7430797Z   skip-logout: false
2025-12-04T08:53:20.7430914Z env:
2025-12-04T08:53:20.7431024Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:20.7431182Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T08:53:20.7431385Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T08:53:20.7431575Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T08:53:20.7432040Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T08:53:20.7432469Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T08:53:20.7432594Z   AWS_REGION: us-east-1
2025-12-04T08:53:20.7432971Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T08:53:20.7433134Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T08:53:20.7435044Z   AWS_SESSION_TOKEN: ***
2025-12-04T08:53:20.7435156Z ##[endgroup]
2025-12-04T08:53:21.1526162Z Logging into registry 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:21.7878645Z ##[group]Run env | grep '^GITHUB' >> "${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}"
2025-12-04T08:53:21.7878996Z env | grep '^GITHUB' >> "${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}"
2025-12-04T08:53:21.7879275Z env | grep '^CI' >> "${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}"
2025-12-04T08:53:21.7879559Z env | grep '^RUNNER' >> "${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}"
2025-12-04T08:53:21.7884915Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:21.7885090Z env:
2025-12-04T08:53:21.7885200Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:21.7885360Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T08:53:21.7885562Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T08:53:21.7885894Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T08:53:21.7886327Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T08:53:21.7886751Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T08:53:21.7886886Z   AWS_REGION: us-east-1
2025-12-04T08:53:21.7887189Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T08:53:21.7887371Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T08:53:21.7889673Z   AWS_SESSION_TOKEN: ***
2025-12-04T08:53:21.7889805Z ##[endgroup]
2025-12-04T08:53:21.8039584Z ##[group]Run pytorch/test-infra/.github/actions/calculate-docker-image@main
2025-12-04T08:53:21.8039767Z with:
2025-12-04T08:53:21.8040049Z   docker-image-name: 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:21.8040365Z   use-custom-docker-registry: true
2025-12-04T08:53:21.8040546Z   docker-build-dir: .ci/docker
2025-12-04T08:53:21.8040674Z   docker-build-script: ./build.sh
2025-12-04T08:53:21.8040801Z   working-directory: .
2025-12-04T08:53:21.8040949Z   docker-registry: 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:21.8041108Z   force-push: false
2025-12-04T08:53:21.8041209Z env:
2025-12-04T08:53:21.8041308Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:21.8041450Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T08:53:21.8041629Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T08:53:21.8041823Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T08:53:21.8042212Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T08:53:21.8042588Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T08:53:21.8042706Z   AWS_REGION: us-east-1
2025-12-04T08:53:21.8042947Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T08:53:21.8043106Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T08:53:21.8045021Z   AWS_SESSION_TOKEN: ***
2025-12-04T08:53:21.8045130Z ##[endgroup]
2025-12-04T08:53:21.8053496Z ##[group]Run set -ex
2025-12-04T08:53:21.8053631Z set -ex
2025-12-04T08:53:21.8053728Z 
2025-12-04T08:53:21.8053886Z # If the docker build directory or the build script doesn't exist, the action will
2025-12-04T08:53:21.8054131Z # gracefully return the docker image name as it is.  Pulling docker image in Linux
2025-12-04T08:53:21.8054349Z # job could then download the pre-built image as usual
2025-12-04T08:53:21.8054606Z if [[ -d "${DOCKER_BUILD_DIR}" ]] && [[ -f "${DOCKER_BUILD_DIR}/${DOCKER_BUILD_SCRIPT}" ]] && [[ "${USE_CUSTOM_DOCKER_REGISTRY}" == "true" ]]; then
2025-12-04T08:53:21.8054846Z   echo "skip=false" >> "${GITHUB_OUTPUT}"
2025-12-04T08:53:21.8054984Z else
2025-12-04T08:53:21.8055098Z   echo "skip=true" >> "${GITHUB_OUTPUT}"
2025-12-04T08:53:21.8055275Z   echo "docker-image=${DOCKER_IMAGE_NAME}" >> "${GITHUB_OUTPUT}"
2025-12-04T08:53:21.8055435Z 
2025-12-04T08:53:21.8055648Z   echo "Not using custom ECR registry.  Either it was not requested or there is no Docker build script in the ${REPO_NAME} repo..."
2025-12-04T08:53:21.8055881Z   exit 0
2025-12-04T08:53:21.8055978Z fi
2025-12-04T08:53:21.8056071Z 
2025-12-04T08:53:21.8056211Z if [[ "${DOCKER_IMAGE_NAME}" == *"${DOCKER_REGISTRY}/${REPO_NAME}"* ]]; then
2025-12-04T08:53:21.8056442Z   # The docker image name already includes the ECR prefix and tag, so we can just
2025-12-04T08:53:21.8056651Z   # use it as it is, but first let's extract the tag
2025-12-04T08:53:21.8056851Z   DOCKER_TAG=$(echo "${DOCKER_IMAGE_NAME}" | awk -F '[:,]' '{print $2}')
2025-12-04T08:53:21.8057119Z   echo "docker-tag=${DOCKER_TAG}" >> "${GITHUB_OUTPUT}"
2025-12-04T08:53:21.8057306Z   echo "docker-image=${DOCKER_IMAGE_NAME}" >> "${GITHUB_OUTPUT}"
2025-12-04T08:53:21.8057462Z else
2025-12-04T08:53:21.8057580Z   if [[ "${DOCKER_IMAGE_NAME}" == *:* ]]; then
2025-12-04T08:53:21.8057735Z     CUSTOM_TAG_PREFIX=${DOCKER_IMAGE_NAME#*:}
2025-12-04T08:53:21.8057892Z     DOCKER_IMAGE_NAME=${DOCKER_IMAGE_NAME%%:*}
2025-12-04T08:53:21.8058028Z   fi
2025-12-04T08:53:21.8058263Z   DOCKER_TAG=${CUSTOM_TAG_PREFIX:+${CUSTOM_TAG_PREFIX}-}$(git rev-parse HEAD:"${DOCKER_BUILD_DIR}")
2025-12-04T08:53:21.8058490Z   echo "docker-tag=${DOCKER_TAG}" >> "${GITHUB_OUTPUT}"
2025-12-04T08:53:21.8058729Z   echo "docker-image=${DOCKER_REGISTRY}/${REPO_NAME}/${DOCKER_IMAGE_NAME}:${DOCKER_TAG}" >> "${GITHUB_OUTPUT}"
2025-12-04T08:53:21.8058986Z   echo "custom-tag-prefix=${CUSTOM_TAG_PREFIX}" >> "${GITHUB_OUTPUT}"
2025-12-04T08:53:21.8059161Z fi
2025-12-04T08:53:21.8061851Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:21.8061998Z env:
2025-12-04T08:53:21.8062097Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:21.8062236Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T08:53:21.8062415Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T08:53:21.8062585Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T08:53:21.8062969Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T08:53:21.8063341Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T08:53:21.8063461Z   AWS_REGION: us-east-1
2025-12-04T08:53:21.8063601Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T08:53:21.8063755Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T08:53:21.8065677Z   AWS_SESSION_TOKEN: ***
2025-12-04T08:53:21.8065786Z   REPO_NAME: pytorch
2025-12-04T08:53:21.8066063Z   DOCKER_IMAGE_NAME: 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:21.8066356Z   DOCKER_BUILD_DIR: .ci/docker
2025-12-04T08:53:21.8066477Z   DOCKER_BUILD_SCRIPT: ./build.sh
2025-12-04T08:53:21.8066631Z   DOCKER_REGISTRY: 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:21.8066790Z   USE_CUSTOM_DOCKER_REGISTRY: true
2025-12-04T08:53:21.8066910Z   CUSTOM_TAG_PREFIX: 
2025-12-04T08:53:21.8067017Z ##[endgroup]
2025-12-04T08:53:21.8082911Z + [[ -d .ci/docker ]]
2025-12-04T08:53:21.8083043Z + [[ -f .ci/docker/./build.sh ]]
2025-12-04T08:53:21.8083170Z + [[ true == \t\r\u\e ]]
2025-12-04T08:53:21.8083286Z + echo skip=false
2025-12-04T08:53:21.8083789Z + [[ 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a == *\3\0\8\5\3\5\3\8\5\1\1\4\.\d\k\r\.\e\c\r\.\u\s\-\e\a\s\t\-\1\.\a\m\a\z\o\n\a\w\s\.\c\o\m\/\p\y\t\o\r\c\h* ]]
2025-12-04T08:53:21.8091011Z ++ echo 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:21.8091303Z ++ awk -F '[:,]' '{print $2}'
2025-12-04T08:53:21.8109176Z + DOCKER_TAG=pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:21.8109876Z + echo docker-tag=pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:21.8110809Z + echo docker-image=308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:21.8139135Z ##[group]Run set +e
2025-12-04T08:53:21.8139335Z set +e
2025-12-04T08:53:21.8139473Z set -x
2025-12-04T08:53:21.8139600Z 
2025-12-04T08:53:21.8139729Z login() {
2025-12-04T08:53:21.8140166Z   aws ecr get-login-password --region us-east-1 | docker login -u AWS --password-stdin "$1"
2025-12-04T08:53:21.8140482Z }
2025-12-04T08:53:21.8140602Z 
2025-12-04T08:53:21.8140723Z retry () {
2025-12-04T08:53:21.8140878Z   $*  || (sleep 1 && $*) || (sleep 2 && $*)
2025-12-04T08:53:21.8141044Z }
2025-12-04T08:53:21.8141179Z 
2025-12-04T08:53:21.8141303Z retry login "${DOCKER_REGISTRY}"
2025-12-04T08:53:21.8141459Z 
2025-12-04T08:53:21.8141577Z START_TIME=$(date +%s)
2025-12-04T08:53:21.8141733Z # Wait up to 120 minutes
2025-12-04T08:53:21.8142103Z while [[ $(( $(date +%s) - 7200 )) -lt $START_TIME ]]; do
2025-12-04T08:53:21.8142351Z   # Check if image already exists, if it does then skip building it
2025-12-04T08:53:21.8142563Z   if docker manifest inspect "${DOCKER_IMAGE}"; then
2025-12-04T08:53:21.8142712Z     exit 0
2025-12-04T08:53:21.8142817Z   fi
2025-12-04T08:53:21.8142908Z 
2025-12-04T08:53:21.8143065Z   # NB: This flag is used by Docker build workflow to push the image to ECR, so we can
2025-12-04T08:53:21.8143321Z   # use this to differentiate between the Docker build and regular build jobs. For the
2025-12-04T08:53:21.8143574Z   # latter, it will wait for the Docker images to become available before continuing
2025-12-04T08:53:21.8143779Z   if [ "${DOCKER_PUSH:-false}" == "true" ]; then
2025-12-04T08:53:21.8143943Z     # It's a Docker build job, let's build the image
2025-12-04T08:53:21.8144082Z     break
2025-12-04T08:53:21.8144185Z   else
2025-12-04T08:53:21.8144324Z     # It's a regular build job, wait for the image to become available
2025-12-04T08:53:21.8144486Z     sleep 300
2025-12-04T08:53:21.8144588Z   fi
2025-12-04T08:53:21.8144680Z done
2025-12-04T08:53:21.8144770Z 
2025-12-04T08:53:21.8144914Z # NB: This part requires a full checkout. Otherwise, the merge base will
2025-12-04T08:53:21.8145135Z # be empty.  The default action would be to continue rebuild the image
2025-12-04T08:53:21.8145339Z if [[ "$BASE_REVISION" = "$(git rev-parse HEAD)" ]]; then
2025-12-04T08:53:21.8145519Z   # if we're on the base branch then use the parent commit
2025-12-04T08:53:21.8145680Z   MERGE_BASE=$(git rev-parse HEAD~)
2025-12-04T08:53:21.8145805Z else
2025-12-04T08:53:21.8145940Z   # otherwise we're on a PR, so use the most recent base commit
2025-12-04T08:53:21.8146132Z   MERGE_BASE=$(git merge-base HEAD "$BASE_REVISION")
2025-12-04T08:53:21.8146274Z fi
2025-12-04T08:53:21.8146365Z 
2025-12-04T08:53:21.8146466Z if [[ -z "${MERGE_BASE}" ]]; then
2025-12-04T08:53:21.8146610Z   echo "rebuild=true" >> "${GITHUB_OUTPUT}"
2025-12-04T08:53:21.8146741Z 
2025-12-04T08:53:21.8146929Z   echo "Finding merge base only works with full checkout, please set fetch-depth to 0, continuing ..."
2025-12-04T08:53:21.8147138Z   exit 0
2025-12-04T08:53:21.8147233Z fi
2025-12-04T08:53:21.8147321Z 
2025-12-04T08:53:21.8147448Z if ! git rev-parse "${MERGE_BASE}:${DOCKER_BUILD_DIR}"; then
2025-12-04T08:53:21.8147713Z   echo "Directory '${DOCKER_BUILD_DIR}' not found in commit $MERGE_BASE, you should rebase onto a more recent commit"
2025-12-04T08:53:21.8147935Z   exit 1
2025-12-04T08:53:21.8148029Z fi
2025-12-04T08:53:21.8148118Z 
2025-12-04T08:53:21.8148268Z PREVIOUS_DOCKER_TAG=$(git rev-parse "${MERGE_BASE}:${DOCKER_BUILD_DIR}")
2025-12-04T08:53:21.8148523Z # If no image exists but the hash is the same as the previous hash then we should error out here
2025-12-04T08:53:21.8148750Z if [[ "${PREVIOUS_DOCKER_TAG}" == "${DOCKER_TAG}" ]]; then
2025-12-04T08:53:21.8149056Z   echo "WARNING: Something has gone wrong and the previous image isn't available for the merge-base of your branch"
2025-12-04T08:53:21.8149344Z   echo "         Will re-build docker image to store in local cache, TTS may be longer"
2025-12-04T08:53:21.8149520Z fi
2025-12-04T08:53:21.8149610Z 
2025-12-04T08:53:21.8149719Z echo "rebuild=true" >> "${GITHUB_OUTPUT}"
2025-12-04T08:53:21.8154107Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:21.8154249Z env:
2025-12-04T08:53:21.8154343Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:21.8154475Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T08:53:21.8154687Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T08:53:21.8154852Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T08:53:21.8155235Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T08:53:21.8155611Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T08:53:21.8155728Z   AWS_REGION: us-east-1
2025-12-04T08:53:21.8155940Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T08:53:21.8156093Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T08:53:21.8158035Z   AWS_SESSION_TOKEN: ***
2025-12-04T08:53:21.8158146Z   DOCKER_BUILD_DIR: .ci/docker
2025-12-04T08:53:21.8158283Z   BASE_REVISION: ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T08:53:21.8158598Z   DOCKER_IMAGE: 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:21.8158953Z   DOCKER_TAG: pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:21.8159181Z   DOCKER_REGISTRY: 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:21.8159332Z   DOCKER_PUSH: 
2025-12-04T08:53:21.8159430Z ##[endgroup]
2025-12-04T08:53:21.8177571Z + retry login 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:21.8178222Z + login 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:21.8180097Z + aws ecr get-login-password --region us-east-1
2025-12-04T08:53:21.8180746Z /home/runner/_work/_temp/54f1b350-2ade-4b0a-8165-e2d3ad7a621f.sh: line 5: aws: command not found
2025-12-04T08:53:21.8181412Z + docker login -u AWS --password-stdin 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:21.8267864Z Error: Cannot perform an interactive login from a non TTY device
2025-12-04T08:53:21.8275273Z + sleep 1
2025-12-04T08:53:22.8285891Z + login 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:22.8289822Z + aws ecr get-login-password --region us-east-1
2025-12-04T08:53:22.8290322Z /home/runner/_work/_temp/54f1b350-2ade-4b0a-8165-e2d3ad7a621f.sh: line 5: aws: command not found
2025-12-04T08:53:22.8291408Z + docker login -u AWS --password-stdin 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:22.8367583Z Error: Cannot perform an interactive login from a non TTY device
2025-12-04T08:53:22.8379781Z + sleep 2
2025-12-04T08:53:24.8393046Z + login 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:24.8396359Z + aws ecr get-login-password --region us-east-1
2025-12-04T08:53:24.8396630Z /home/runner/_work/_temp/54f1b350-2ade-4b0a-8165-e2d3ad7a621f.sh: line 5: aws: command not found
2025-12-04T08:53:24.8397926Z + docker login -u AWS --password-stdin 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:24.8488314Z Error: Cannot perform an interactive login from a non TTY device
2025-12-04T08:53:24.8499825Z ++ date +%s
2025-12-04T08:53:24.8509327Z + START_TIME=1764838404
2025-12-04T08:53:24.8512894Z ++ date +%s
2025-12-04T08:53:24.8519739Z + [[ 1764831204 -lt 1764838404 ]]
2025-12-04T08:53:24.8520389Z + docker manifest inspect 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:26.2101027Z {
2025-12-04T08:53:26.2101379Z 	"schemaVersion": 2,
2025-12-04T08:53:26.2101904Z 	"mediaType": "application/vnd.docker.distribution.manifest.v2+json",
2025-12-04T08:53:26.2102401Z 	"config": {
2025-12-04T08:53:26.2102765Z 		"mediaType": "application/vnd.docker.container.image.v1+json",
2025-12-04T08:53:26.2103190Z 		"size": 30522,
2025-12-04T08:53:26.2103637Z 		"digest": "sha256:79498ef00fdf8abfcde955fd685c3a7412c33ca80383b5905abfdc3c70621215"
2025-12-04T08:53:26.2104128Z 	},
2025-12-04T08:53:26.2104348Z 	"layers": [
2025-12-04T08:53:26.2104575Z 		{
2025-12-04T08:53:26.2104930Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2105371Z 			"size": 30594402,
2025-12-04T08:53:26.2106623Z 			"digest": "sha256:02de03a7213b62b792ec66a7efb8c86c4117ca00fb8651facf8ecfe33044b485"
2025-12-04T08:53:26.2107109Z 		},
2025-12-04T08:53:26.2107321Z 		{
2025-12-04T08:53:26.2107675Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2108093Z 			"size": 1554,
2025-12-04T08:53:26.2108537Z 			"digest": "sha256:3a5718b5258e28918133dd74ea64bd506b2c15530a2fa8a72c45c5b0d8f7c7b0"
2025-12-04T08:53:26.2109004Z 		},
2025-12-04T08:53:26.2109212Z 		{
2025-12-04T08:53:26.2109548Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2109977Z 			"size": 335779211,
2025-12-04T08:53:26.2110497Z 			"digest": "sha256:bf3aa22776924a41b55849f0f30cb22af45d41da1177a9d682cf94cde99d8f98"
2025-12-04T08:53:26.2110966Z 		},
2025-12-04T08:53:26.2111175Z 		{
2025-12-04T08:53:26.2111517Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2111929Z 			"size": 704,
2025-12-04T08:53:26.2112320Z 			"digest": "sha256:9d58e5257cefd43e8226153d71d28a865253662146aa9fce9a9f95af67b497fa"
2025-12-04T08:53:26.2124872Z 		},
2025-12-04T08:53:26.2125000Z 		{
2025-12-04T08:53:26.2125321Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2125562Z 			"size": 1770,
2025-12-04T08:53:26.2125759Z 			"digest": "sha256:fde80a64553533a56c032d4bc388837e7d4631a0424d1bfe135703165b67fd4d"
2025-12-04T08:53:26.2126048Z 		},
2025-12-04T08:53:26.2126215Z 		{
2025-12-04T08:53:26.2126379Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2126561Z 			"size": 485,
2025-12-04T08:53:26.2126741Z 			"digest": "sha256:6931c5f20e80e481e4f484471ff3a02878b4f8c54a9a5a4717213fdaa35c0bff"
2025-12-04T08:53:26.2126939Z 		},
2025-12-04T08:53:26.2127027Z 		{
2025-12-04T08:53:26.2127172Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2127354Z 			"size": 120663474,
2025-12-04T08:53:26.2127551Z 			"digest": "sha256:170ea6d3edd62991e37d2e6ebe53dfcd4601f5d42e8f9720af5f8db5fc267856"
2025-12-04T08:53:26.2127753Z 		},
2025-12-04T08:53:26.2127844Z 		{
2025-12-04T08:53:26.2127989Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2128167Z 			"size": 4433,
2025-12-04T08:53:26.2128349Z 			"digest": "sha256:dc8487f6c81cac00fa33031f8d3481e2c3634c4f064a9c4c36b87b41e78bc9fb"
2025-12-04T08:53:26.2128551Z 		},
2025-12-04T08:53:26.2128641Z 		{
2025-12-04T08:53:26.2128785Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2128964Z 			"size": 1755,
2025-12-04T08:53:26.2129142Z 			"digest": "sha256:9748c5348f39a11c960c49fd9219fdea1c23e612ed11a02d71501424defc80f5"
2025-12-04T08:53:26.2129341Z 		},
2025-12-04T08:53:26.2129433Z 		{
2025-12-04T08:53:26.2129576Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2129751Z 			"size": 724,
2025-12-04T08:53:26.2129932Z 			"digest": "sha256:8539cc3f8d8a138501ed0255c0cd7ec491bc0add9e4a62095f1c0f9533daa1cc"
2025-12-04T08:53:26.2130135Z 		},
2025-12-04T08:53:26.2130226Z 		{
2025-12-04T08:53:26.2130370Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2130621Z 			"size": 3378352584,
2025-12-04T08:53:26.2130812Z 			"digest": "sha256:af88f886884fe6f1a1992efb7ce8473901f795eef69caa199443f3e076fdfd5b"
2025-12-04T08:53:26.2131153Z 		},
2025-12-04T08:53:26.2131243Z 		{
2025-12-04T08:53:26.2131387Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2131564Z 			"size": 396,
2025-12-04T08:53:26.2131743Z 			"digest": "sha256:32fbb88555c4195c45c7008cf92e389d67acc79a7e382503003ef93bcb886afe"
2025-12-04T08:53:26.2131939Z 		},
2025-12-04T08:53:26.2132028Z 		{
2025-12-04T08:53:26.2132169Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2132324Z 			"size": 80171601,
2025-12-04T08:53:26.2132480Z 			"digest": "sha256:3231e1ab814b143b244037c540b637be259085834865ac43b1ed2b6f6ad631e1"
2025-12-04T08:53:26.2132648Z 		},
2025-12-04T08:53:26.2132777Z 		{
2025-12-04T08:53:26.2132902Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2133057Z 			"size": 787,
2025-12-04T08:53:26.2133214Z 			"digest": "sha256:80061bf5dcbb9a4e38ac865a9cdc0a615bb294e3e6bfa357a6d515dcf3f54abc"
2025-12-04T08:53:26.2133391Z 		},
2025-12-04T08:53:26.2133473Z 		{
2025-12-04T08:53:26.2133599Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2133752Z 			"size": 106,
2025-12-04T08:53:26.2133907Z 			"digest": "sha256:6e9524f4518ec02b47ff12c55b6b6afbc65b3f4be59072e2afe20c2c87522549"
2025-12-04T08:53:26.2134080Z 		},
2025-12-04T08:53:26.2134158Z 		{
2025-12-04T08:53:26.2134283Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2134436Z 			"size": 1495,
2025-12-04T08:53:26.2134591Z 			"digest": "sha256:ce919d4bf5eeff71d49b160a16603117225530497c3905e02224227d11e2ff88"
2025-12-04T08:53:26.2134762Z 		},
2025-12-04T08:53:26.2134839Z 		{
2025-12-04T08:53:26.2134969Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2135125Z 			"size": 548601195,
2025-12-04T08:53:26.2135284Z 			"digest": "sha256:47681e3e6f37423139a5c86549ffbb43e4f258344b0461208f5821263da152e9"
2025-12-04T08:53:26.2135453Z 		},
2025-12-04T08:53:26.2135532Z 		{
2025-12-04T08:53:26.2135663Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2135817Z 			"size": 162,
2025-12-04T08:53:26.2135974Z 			"digest": "sha256:cb70fe22c9ebacebfe8402519059c8a66da6d5a77979e4c0ecdb3a762bebe357"
2025-12-04T08:53:26.2136149Z 		},
2025-12-04T08:53:26.2136228Z 		{
2025-12-04T08:53:26.2136352Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2136506Z 			"size": 104,
2025-12-04T08:53:26.2136662Z 			"digest": "sha256:17858e829c8cfe9a7e22516e03ad5273d8cf5c50f58edb10ff60c74e15c8e1f6"
2025-12-04T08:53:26.2136836Z 		},
2025-12-04T08:53:26.2136912Z 		{
2025-12-04T08:53:26.2137037Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2137195Z 			"size": 724,
2025-12-04T08:53:26.2137352Z 			"digest": "sha256:8539cc3f8d8a138501ed0255c0cd7ec491bc0add9e4a62095f1c0f9533daa1cc"
2025-12-04T08:53:26.2137527Z 		},
2025-12-04T08:53:26.2137608Z 		{
2025-12-04T08:53:26.2137733Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2137887Z 			"size": 196,
2025-12-04T08:53:26.2138041Z 			"digest": "sha256:a63f3b4eed1157bcb3c51b64196e74e9f10d1f923652b02fd433c6ed993597ff"
2025-12-04T08:53:26.2138216Z 		},
2025-12-04T08:53:26.2138298Z 		{
2025-12-04T08:53:26.2138421Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2138572Z 			"size": 2584,
2025-12-04T08:53:26.2138731Z 			"digest": "sha256:10ab3d1afbc4cb2d3ced8f3e0072c0b1dd124dcadcf68b95fadf8a7a9f663860"
2025-12-04T08:53:26.2138907Z 		},
2025-12-04T08:53:26.2138983Z 		{
2025-12-04T08:53:26.2139105Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2139262Z 			"size": 7652105336,
2025-12-04T08:53:26.2139421Z 			"digest": "sha256:98ca88b5095b449a2f2d753a21217856271912fbe51c2d99f928a2196f4097d5"
2025-12-04T08:53:26.2139593Z 		},
2025-12-04T08:53:26.2139670Z 		{
2025-12-04T08:53:26.2139792Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2140012Z 			"size": 135,
2025-12-04T08:53:26.2140163Z 			"digest": "sha256:025c90839a58c768b3cc444e48cae67c1a5b2c85320ad8827231f0ba390cf9aa"
2025-12-04T08:53:26.2140332Z 		},
2025-12-04T08:53:26.2140448Z 		{
2025-12-04T08:53:26.2140575Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2140725Z 			"size": 104,
2025-12-04T08:53:26.2140877Z 			"digest": "sha256:9255df5942ae69fee24f8074314f451d5d2f1ca71b6c777274297fd43a0032d8"
2025-12-04T08:53:26.2141045Z 		},
2025-12-04T08:53:26.2141121Z 		{
2025-12-04T08:53:26.2141243Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2141394Z 			"size": 612,
2025-12-04T08:53:26.2141597Z 			"digest": "sha256:f71ca9d4ed1c4ca8177602f3cb0db83d9787ea6c258a8ef203387b308ff3e0f0"
2025-12-04T08:53:26.2141768Z 		},
2025-12-04T08:53:26.2141842Z 		{
2025-12-04T08:53:26.2141963Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2142117Z 			"size": 838191953,
2025-12-04T08:53:26.2142282Z 			"digest": "sha256:d02b47b56ca7f3598f5943d4fdc7139d5e3d3bc82d49185cedf9817dd55fc75c"
2025-12-04T08:53:26.2142454Z 		},
2025-12-04T08:53:26.2142530Z 		{
2025-12-04T08:53:26.2142653Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2142805Z 			"size": 111,
2025-12-04T08:53:26.2142956Z 			"digest": "sha256:40279492aea7bc8fb650842b495912195621c21b14cef4c717a9e0a9fc535131"
2025-12-04T08:53:26.2143122Z 		},
2025-12-04T08:53:26.2143200Z 		{
2025-12-04T08:53:26.2143322Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2143472Z 			"size": 1556,
2025-12-04T08:53:26.2143630Z 			"digest": "sha256:33a27ce74abd7e32a03a564fc45005bc75904b53ad516f18d47facbeb2f2794e"
2025-12-04T08:53:26.2143799Z 		},
2025-12-04T08:53:26.2143874Z 		{
2025-12-04T08:53:26.2143995Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2144144Z 			"size": 107,
2025-12-04T08:53:26.2144302Z 			"digest": "sha256:6b66ed335d1d8df6140caba76d9c2babed83bb37962e1e638825d49e67184fa5"
2025-12-04T08:53:26.2144473Z 		},
2025-12-04T08:53:26.2144551Z 		{
2025-12-04T08:53:26.2144674Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2144823Z 			"size": 166,
2025-12-04T08:53:26.2144976Z 			"digest": "sha256:9f010fa04118bfee2d7b4481e6badb714032bde0652b04151a6599e57e1bd91b"
2025-12-04T08:53:26.2145147Z 		},
2025-12-04T08:53:26.2145225Z 		{
2025-12-04T08:53:26.2145347Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2145499Z 			"size": 3702493,
2025-12-04T08:53:26.2145662Z 			"digest": "sha256:6c64d5e8bb6ae6ef4e3f1d316429d8b14a6e8a1fb410fb83b96c8bbd4a0a095c"
2025-12-04T08:53:26.2145835Z 		},
2025-12-04T08:53:26.2145911Z 		{
2025-12-04T08:53:26.2146033Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2146187Z 			"size": 107,
2025-12-04T08:53:26.2146339Z 			"digest": "sha256:c20ea058f549f5f5538c95c5e0da23afbbc9fb7ffc1987d126fe684eeed743f5"
2025-12-04T08:53:26.2146513Z 		},
2025-12-04T08:53:26.2146588Z 		{
2025-12-04T08:53:26.2146711Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2146863Z 			"size": 829,
2025-12-04T08:53:26.2147014Z 			"digest": "sha256:3c4fd2d54638a1336d39769fe36041aa6d186a8dea0e7096b8d8a7068ba0d3c0"
2025-12-04T08:53:26.2147181Z 		},
2025-12-04T08:53:26.2147257Z 		{
2025-12-04T08:53:26.2147380Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2147533Z 			"size": 26673844,
2025-12-04T08:53:26.2147694Z 			"digest": "sha256:964ebac3d7a95c64ea7f0d828cd58e6244cc955e9a099a2525079ecf64026e3f"
2025-12-04T08:53:26.2147867Z 		},
2025-12-04T08:53:26.2147942Z 		{
2025-12-04T08:53:26.2148066Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2148218Z 			"size": 104,
2025-12-04T08:53:26.2148372Z 			"digest": "sha256:2aaa7210673fc5bd15d36e54ee5c3fb495d1eafa1cb8d686054ccedb1c37bfc8"
2025-12-04T08:53:26.2148587Z 		},
2025-12-04T08:53:26.2148665Z 		{
2025-12-04T08:53:26.2148786Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2148939Z 			"size": 424,
2025-12-04T08:53:26.2149092Z 			"digest": "sha256:fa273daa00371a98ed668535e14b8cc3cb425feba0b601b3e3c72314d0234312"
2025-12-04T08:53:26.2149264Z 		},
2025-12-04T08:53:26.2149341Z 		{
2025-12-04T08:53:26.2149465Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2149618Z 			"size": 19279582,
2025-12-04T08:53:26.2149779Z 			"digest": "sha256:d931a62fd2408369decfa0e6eac11768e35d0ffddee87d769c82aaf1ad7e2899"
2025-12-04T08:53:26.2149954Z 		},
2025-12-04T08:53:26.2150066Z 		{
2025-12-04T08:53:26.2150189Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2150343Z 			"size": 826,
2025-12-04T08:53:26.2150533Z 			"digest": "sha256:d3573d61c28e1400840260d3c2c786c9e104f6558162beac799e55b6f5c1e747"
2025-12-04T08:53:26.2150703Z 		},
2025-12-04T08:53:26.2150785Z 		{
2025-12-04T08:53:26.2150904Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2151058Z 			"size": 724,
2025-12-04T08:53:26.2151212Z 			"digest": "sha256:8539cc3f8d8a138501ed0255c0cd7ec491bc0add9e4a62095f1c0f9533daa1cc"
2025-12-04T08:53:26.2151383Z 		},
2025-12-04T08:53:26.2151461Z 		{
2025-12-04T08:53:26.2151583Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2151734Z 			"size": 149,
2025-12-04T08:53:26.2151888Z 			"digest": "sha256:f9b32f08c49055dd61bd359d5f42f6adb9e5a183c2821d97d11572dd7ce1e91f"
2025-12-04T08:53:26.2152058Z 		},
2025-12-04T08:53:26.2152135Z 		{
2025-12-04T08:53:26.2152261Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2152412Z 			"size": 136,
2025-12-04T08:53:26.2152562Z 			"digest": "sha256:3a0206399d60f6e8897f78c8e8f81b59d51969a329ef45485d28ae19607ca72c"
2025-12-04T08:53:26.2152730Z 		},
2025-12-04T08:53:26.2152807Z 		{
2025-12-04T08:53:26.2152933Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2153084Z 			"size": 140,
2025-12-04T08:53:26.2153235Z 			"digest": "sha256:386f322edd1c1c275126bab065c22fcd3950916c1fb8491a21a7f5c358af599a"
2025-12-04T08:53:26.2153404Z 		},
2025-12-04T08:53:26.2153480Z 		{
2025-12-04T08:53:26.2153603Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2153756Z 			"size": 32,
2025-12-04T08:53:26.2153911Z 			"digest": "sha256:4f4fb700ef54461cfa02571ae0db9a0dc1e0cdb5577484a6d75e68dc38e8acc1"
2025-12-04T08:53:26.2154084Z 		},
2025-12-04T08:53:26.2154161Z 		{
2025-12-04T08:53:26.2154287Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2154438Z 			"size": 223,
2025-12-04T08:53:26.2154591Z 			"digest": "sha256:bbe49df30697f6959cd958299909d9255cd54663ce2e9e2c2d378f8f9dfe8345"
2025-12-04T08:53:26.2154762Z 		},
2025-12-04T08:53:26.2154841Z 		{
2025-12-04T08:53:26.2154965Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2155118Z 			"size": 346,
2025-12-04T08:53:26.2155272Z 			"digest": "sha256:d6630aa6f375b12cb7471c5b60eb32e02ff8d70adf4497e061d6c15fead186c7"
2025-12-04T08:53:26.2155442Z 		},
2025-12-04T08:53:26.2155516Z 		{
2025-12-04T08:53:26.2155640Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2155793Z 			"size": 88302,
2025-12-04T08:53:26.2155951Z 			"digest": "sha256:6d807afc1309592c99c7d77af3874afb54c1718377fe721ac0cc616f59d291b9"
2025-12-04T08:53:26.2156119Z 		},
2025-12-04T08:53:26.2156195Z 		{
2025-12-04T08:53:26.2156318Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2156471Z 			"size": 106,
2025-12-04T08:53:26.2156621Z 			"digest": "sha256:60b679430e4e0b7690392dfe4f5dc417847f7a3ba2b761ce747b66d412e1d956"
2025-12-04T08:53:26.2156790Z 		},
2025-12-04T08:53:26.2156868Z 		{
2025-12-04T08:53:26.2156990Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2157181Z 			"size": 1671,
2025-12-04T08:53:26.2157338Z 			"digest": "sha256:3992ae84f9eda1c5c52fa96b1f1d0fc3f93c661c5cf0b971a504a260c290da49"
2025-12-04T08:53:26.2157509Z 		},
2025-12-04T08:53:26.2157587Z 		{
2025-12-04T08:53:26.2157710Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2157862Z 			"size": 724,
2025-12-04T08:53:26.2158012Z 			"digest": "sha256:8539cc3f8d8a138501ed0255c0cd7ec491bc0add9e4a62095f1c0f9533daa1cc"
2025-12-04T08:53:26.2158181Z 		},
2025-12-04T08:53:26.2158260Z 		{
2025-12-04T08:53:26.2158385Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2158536Z 			"size": 138,
2025-12-04T08:53:26.2158724Z 			"digest": "sha256:62d400609f9c38fce4745f72372423072ba0f142b3c03775ccb317f6c5240966"
2025-12-04T08:53:26.2158891Z 		},
2025-12-04T08:53:26.2158967Z 		{
2025-12-04T08:53:26.2159090Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2159241Z 			"size": 119,
2025-12-04T08:53:26.2159396Z 			"digest": "sha256:7e7b097490967d568331cc9f8afdd02422fe101c6364ec5e12dba2970991e533"
2025-12-04T08:53:26.2159561Z 		},
2025-12-04T08:53:26.2159641Z 		{
2025-12-04T08:53:26.2159764Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2159918Z 			"size": 6231259865,
2025-12-04T08:53:26.2160083Z 			"digest": "sha256:7dcdbd8421cb17aaa5d0cb965ddf94e196cb364e762b12ab78024cb25e3b6bcd"
2025-12-04T08:53:26.2160256Z 		},
2025-12-04T08:53:26.2160334Z 		{
2025-12-04T08:53:26.2160509Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2160662Z 			"size": 174,
2025-12-04T08:53:26.2160817Z 			"digest": "sha256:cbb12613719bab9f179968227f9fb8881251992804e460b9a9e1c00f3ac4a0c5"
2025-12-04T08:53:26.2160985Z 		},
2025-12-04T08:53:26.2161063Z 		{
2025-12-04T08:53:26.2161187Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2161339Z 			"size": 1896,
2025-12-04T08:53:26.2161497Z 			"digest": "sha256:e87038dce9bc8e13bd64006847d30ddcaf77455256c4985fccfec83f82d4b925"
2025-12-04T08:53:26.2161672Z 		},
2025-12-04T08:53:26.2161750Z 		{
2025-12-04T08:53:26.2161876Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2162030Z 			"size": 162783968,
2025-12-04T08:53:26.2162190Z 			"digest": "sha256:e4606b636f96f1c80f4be26aeb9d6f5f990f6149789c2de160451c5ac76a467d"
2025-12-04T08:53:26.2162360Z 		},
2025-12-04T08:53:26.2162437Z 		{
2025-12-04T08:53:26.2162559Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2162711Z 			"size": 302,
2025-12-04T08:53:26.2162864Z 			"digest": "sha256:6f2a5d33b946e561219b9968769773e36ce1d28bee8c62eff652098b7825fc79"
2025-12-04T08:53:26.2163032Z 		},
2025-12-04T08:53:26.2163108Z 		{
2025-12-04T08:53:26.2163231Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2163384Z 			"size": 32,
2025-12-04T08:53:26.2163540Z 			"digest": "sha256:4f4fb700ef54461cfa02571ae0db9a0dc1e0cdb5577484a6d75e68dc38e8acc1"
2025-12-04T08:53:26.2163715Z 		},
2025-12-04T08:53:26.2163791Z 		{
2025-12-04T08:53:26.2163913Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2164065Z 			"size": 108,
2025-12-04T08:53:26.2164219Z 			"digest": "sha256:a4f2bf2f19e63b91d46f2d9cf11a25c657517a6835996404da1e79a09d918b0e"
2025-12-04T08:53:26.2164388Z 		},
2025-12-04T08:53:26.2164461Z 		{
2025-12-04T08:53:26.2164582Z 			"mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip",
2025-12-04T08:53:26.2164732Z 			"size": 54145661,
2025-12-04T08:53:26.2164892Z 			"digest": "sha256:1ae00acdac56cbc6d3f81b3c5d854a2b77c30d458b0fbe18c5935145364484f0"
2025-12-04T08:53:26.2165066Z 		}
2025-12-04T08:53:26.2165142Z 	]
2025-12-04T08:53:26.2165222Z }
2025-12-04T08:53:26.2165320Z + exit 0
2025-12-04T08:53:26.2180522Z ##[group]Run set -eux
2025-12-04T08:53:26.2180638Z set -eux
2025-12-04T08:53:26.2180799Z # It's ok if this steps fails, it would then be an anonymous user like what we used to have
2025-12-04T08:53:26.2181268Z aws secretsmanager get-secret-value --secret-id docker_hub_readonly_token | jq --raw-output '.SecretString' | jq -r .docker_hub_readonly_token | docker login --username pytorchbot --password-stdin || true
2025-12-04T08:53:26.2185525Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:26.2185679Z env:
2025-12-04T08:53:26.2185772Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:26.2185906Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T08:53:26.2186086Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T08:53:26.2186253Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T08:53:26.2186699Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T08:53:26.2187072Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T08:53:26.2187194Z   AWS_REGION: us-east-1
2025-12-04T08:53:26.2187378Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T08:53:26.2187535Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T08:53:26.2189511Z   AWS_SESSION_TOKEN: ***
2025-12-04T08:53:26.2189617Z ##[endgroup]
2025-12-04T08:53:26.2213641Z + aws secretsmanager get-secret-value --secret-id docker_hub_readonly_token
2025-12-04T08:53:26.2214082Z + jq --raw-output .SecretString
2025-12-04T08:53:26.2214491Z /home/runner/_work/_temp/f1bda5c4-c1f1-43ab-a43d-2a7889e6484a.sh: line 3: aws: command not found
2025-12-04T08:53:26.2215668Z + jq -r .docker_hub_readonly_token
2025-12-04T08:53:26.2216122Z + docker login --username pytorchbot --password-stdin
2025-12-04T08:53:26.2309704Z Error: Cannot perform an interactive login from a non TTY device
2025-12-04T08:53:26.2315928Z + true
2025-12-04T08:53:26.2375483Z ##[group]Run pytorch/test-infra/.github/actions/pull-docker-image@main
2025-12-04T08:53:26.2375703Z with:
2025-12-04T08:53:26.2375972Z   docker-image: 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:26.2376326Z   docker-registry: 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:26.2376491Z env:
2025-12-04T08:53:26.2376585Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:26.2376719Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T08:53:26.2376892Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T08:53:26.2377058Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T08:53:26.2377445Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T08:53:26.2377853Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T08:53:26.2377973Z   AWS_REGION: us-east-1
2025-12-04T08:53:26.2378213Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T08:53:26.2378375Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T08:53:26.2380314Z   AWS_SESSION_TOKEN: ***
2025-12-04T08:53:26.2380474Z ##[endgroup]
2025-12-04T08:53:26.2387129Z ##[group]Run set -x
2025-12-04T08:53:26.2387250Z set -x
2025-12-04T08:53:26.2387347Z set +e
2025-12-04T08:53:26.2387443Z 
2025-12-04T08:53:26.2387530Z login() {
2025-12-04T08:53:26.2387717Z   aws ecr get-login-password --region us-east-1 | docker login -u AWS --password-stdin "$1"
2025-12-04T08:53:26.2387913Z }
2025-12-04T08:53:26.2387998Z 
2025-12-04T08:53:26.2388087Z retry () {
2025-12-04T08:53:26.2388198Z   $*  || (sleep 1 && $*) || (sleep 2 && $*)
2025-12-04T08:53:26.2388322Z }
2025-12-04T08:53:26.2388407Z 
2025-12-04T08:53:26.2388506Z retry login "${DOCKER_REGISTRY}"
2025-12-04T08:53:26.2388624Z 
2025-12-04T08:53:26.2388806Z IMAGE_SIZE=$(docker manifest inspect "${DOCKER_IMAGE}" | jq '[.layers[].size, .config.size] | add / 1024 / 1024')
2025-12-04T08:53:26.2389185Z echo "Compressed size of image in MB: ${IMAGE_SIZE}"
2025-12-04T08:53:26.2389327Z 
2025-12-04T08:53:26.2389412Z set -e
2025-12-04T08:53:26.2389554Z # ignore output since only exit code is used for conditional
2025-12-04T08:53:26.2389740Z # only pull docker image if it's not available locally
2025-12-04T08:53:26.2389946Z if ! docker inspect --type=image "${DOCKER_IMAGE}" >/dev/null 2>/dev/null; then
2025-12-04T08:53:26.2390135Z   retry docker pull "${DOCKER_IMAGE}"
2025-12-04T08:53:26.2390260Z fi
2025-12-04T08:53:26.2394522Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T08:53:26.2394670Z env:
2025-12-04T08:53:26.2394768Z   GIT_DEFAULT_BRANCH: main
2025-12-04T08:53:26.2394912Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T08:53:26.2395092Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T08:53:26.2395270Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T08:53:26.2395658Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T08:53:26.2396035Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T08:53:26.2396156Z   AWS_REGION: us-east-1
2025-12-04T08:53:26.2396296Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T08:53:26.2396452Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T08:53:26.2398397Z   AWS_SESSION_TOKEN: ***
2025-12-04T08:53:26.2398678Z   DOCKER_IMAGE: 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:26.2399082Z   DOCKER_REGISTRY: 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:26.2399236Z ##[endgroup]
2025-12-04T08:53:26.2418659Z + set +e
2025-12-04T08:53:26.2418903Z + retry login 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:26.2419238Z + login 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:26.2422331Z + aws ecr get-login-password --region us-east-1
2025-12-04T08:53:26.2422716Z /home/runner/_work/_temp/862a5edd-ff9a-4a54-84fc-714af04c51e3.sh: line 5: aws: command not found
2025-12-04T08:53:26.2423478Z + docker login -u AWS --password-stdin 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:26.2507950Z Error: Cannot perform an interactive login from a non TTY device
2025-12-04T08:53:26.2515958Z + sleep 1
2025-12-04T08:53:27.2530904Z + login 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:27.2535733Z + aws ecr get-login-password --region us-east-1
2025-12-04T08:53:27.2537056Z /home/runner/_work/_temp/862a5edd-ff9a-4a54-84fc-714af04c51e3.sh: line 5: aws: command not found
2025-12-04T08:53:27.2537767Z + docker login -u AWS --password-stdin 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:27.2618110Z Error: Cannot perform an interactive login from a non TTY device
2025-12-04T08:53:27.2627926Z + sleep 2
2025-12-04T08:53:29.2641531Z + login 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:29.2645084Z + aws ecr get-login-password --region us-east-1
2025-12-04T08:53:29.2645753Z /home/runner/_work/_temp/862a5edd-ff9a-4a54-84fc-714af04c51e3.sh: line 5: aws: command not found
2025-12-04T08:53:29.2646500Z + docker login -u AWS --password-stdin 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T08:53:29.2737620Z Error: Cannot perform an interactive login from a non TTY device
2025-12-04T08:53:29.2752684Z ++ docker manifest inspect 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:29.2753285Z ++ jq '[.layers[].size, .config.size] | add / 1024 / 1024'
2025-12-04T08:53:30.6379555Z + IMAGE_SIZE=18579.916069984436
2025-12-04T08:53:30.6379849Z + echo 'Compressed size of image in MB: 18579.916069984436'
2025-12-04T08:53:30.6380084Z + set -e
2025-12-04T08:53:30.6380543Z Compressed size of image in MB: 18579.916069984436
2025-12-04T08:53:30.6381059Z + docker inspect --type=image 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:30.6487208Z + retry docker pull 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:30.6487858Z + docker pull 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T08:53:31.7063029Z pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a: Pulling from pytorch/ci-image
2025-12-04T08:53:31.7063563Z 02de03a7213b: Pulling fs layer
2025-12-04T08:53:31.7063813Z 3a5718b5258e: Pulling fs layer
2025-12-04T08:53:31.7064032Z bf3aa2277692: Pulling fs layer
2025-12-04T08:53:31.7064233Z 9d58e5257cef: Pulling fs layer
2025-12-04T08:53:31.7064445Z fde80a645535: Pulling fs layer
2025-12-04T08:53:31.7081572Z 6931c5f20e80: Pulling fs layer
2025-12-04T08:53:31.7081780Z 170ea6d3edd6: Pulling fs layer
2025-12-04T08:53:31.7081993Z dc8487f6c81c: Pulling fs layer
2025-12-04T08:53:31.7082177Z 9748c5348f39: Pulling fs layer
2025-12-04T08:53:31.7082341Z 8539cc3f8d8a: Pulling fs layer
2025-12-04T08:53:31.7082500Z af88f886884f: Pulling fs layer
2025-12-04T08:53:31.7082725Z 32fbb88555c4: Pulling fs layer
2025-12-04T08:53:31.7082897Z 3231e1ab814b: Pulling fs layer
2025-12-04T08:53:31.7083061Z 80061bf5dcbb: Pulling fs layer
2025-12-04T08:53:31.7083220Z 6e9524f4518e: Pulling fs layer
2025-12-04T08:53:31.7083380Z ce919d4bf5ee: Pulling fs layer
2025-12-04T08:53:31.7083539Z 47681e3e6f37: Pulling fs layer
2025-12-04T08:53:31.7083698Z cb70fe22c9eb: Pulling fs layer
2025-12-04T08:53:31.7084231Z 17858e829c8c: Pulling fs layer
2025-12-04T08:53:31.7084393Z a63f3b4eed11: Pulling fs layer
2025-12-04T08:53:31.7084556Z 10ab3d1afbc4: Pulling fs layer
2025-12-04T08:53:31.7084721Z 98ca88b5095b: Pulling fs layer
2025-12-04T08:53:31.7084879Z 025c90839a58: Pulling fs layer
2025-12-04T08:53:31.7085037Z 9255df5942ae: Pulling fs layer
2025-12-04T08:53:31.7095853Z f71ca9d4ed1c: Pulling fs layer
2025-12-04T08:53:31.7096006Z 9d58e5257cef: Waiting
2025-12-04T08:53:31.7096146Z d02b47b56ca7: Pulling fs layer
2025-12-04T08:53:31.7096287Z 40279492aea7: Pulling fs layer
2025-12-04T08:53:31.7096438Z fde80a645535: Waiting
2025-12-04T08:53:31.7096573Z 33a27ce74abd: Pulling fs layer
2025-12-04T08:53:31.7096714Z 6b66ed335d1d: Pulling fs layer
2025-12-04T08:53:31.7096856Z 9f010fa04118: Pulling fs layer
2025-12-04T08:53:31.7096996Z 6c64d5e8bb6a: Pulling fs layer
2025-12-04T08:53:31.7097128Z 8539cc3f8d8a: Waiting
2025-12-04T08:53:31.7097253Z c20ea058f549: Pulling fs layer
2025-12-04T08:53:31.7097442Z 3c4fd2d54638: Pulling fs layer
2025-12-04T08:53:31.7097577Z af88f886884f: Waiting
2025-12-04T08:53:31.7097698Z 32fbb88555c4: Waiting
2025-12-04T08:53:31.7097818Z 80061bf5dcbb: Waiting
2025-12-04T08:53:31.7097941Z 3231e1ab814b: Waiting
2025-12-04T08:53:31.7098059Z dc8487f6c81c: Waiting
2025-12-04T08:53:31.7098179Z 6e9524f4518e: Waiting
2025-12-04T08:53:31.7098300Z a63f3b4eed11: Waiting
2025-12-04T08:53:31.7098415Z 9748c5348f39: Waiting
2025-12-04T08:53:31.7098531Z 025c90839a58: Waiting
2025-12-04T08:53:31.7098648Z 98ca88b5095b: Waiting
2025-12-04T08:53:31.7098770Z 9255df5942ae: Waiting
2025-12-04T08:53:31.7098891Z 10ab3d1afbc4: Waiting
2025-12-04T08:53:31.7099008Z d02b47b56ca7: Waiting
2025-12-04T08:53:31.7099124Z 40279492aea7: Waiting
2025-12-04T08:53:31.7099240Z 6c64d5e8bb6a: Waiting
2025-12-04T08:53:31.7099359Z c20ea058f549: Waiting
2025-12-04T08:53:31.7099478Z ce919d4bf5ee: Waiting
2025-12-04T08:53:31.7099595Z 6b66ed335d1d: Waiting
2025-12-04T08:53:31.7099711Z 47681e3e6f37: Waiting
2025-12-04T08:53:31.7099832Z 170ea6d3edd6: Waiting
2025-12-04T08:53:31.7099949Z cb70fe22c9eb: Waiting
2025-12-04T08:53:31.7100066Z 17858e829c8c: Waiting
2025-12-04T08:53:31.7100182Z 6931c5f20e80: Waiting
2025-12-04T08:53:31.7100536Z 964ebac3d7a9: Pulling fs layer
2025-12-04T08:53:31.7100677Z 2aaa7210673f: Pulling fs layer
2025-12-04T08:53:31.7100811Z 3c4fd2d54638: Waiting
2025-12-04T08:53:31.7100939Z fa273daa0037: Pulling fs layer
2025-12-04T08:53:31.7101074Z 964ebac3d7a9: Waiting
2025-12-04T08:53:31.7101192Z 2aaa7210673f: Waiting
2025-12-04T08:53:31.7101317Z d931a62fd240: Pulling fs layer
2025-12-04T08:53:31.7101456Z d3573d61c28e: Pulling fs layer
2025-12-04T08:53:31.7101588Z fa273daa0037: Waiting
2025-12-04T08:53:31.7101706Z d931a62fd240: Waiting
2025-12-04T08:53:31.7101824Z d3573d61c28e: Waiting
2025-12-04T08:53:31.7101948Z f9b32f08c490: Pulling fs layer
2025-12-04T08:53:31.7102088Z 3a0206399d60: Pulling fs layer
2025-12-04T08:53:31.7102221Z f9b32f08c490: Waiting
2025-12-04T08:53:31.7102338Z 3a0206399d60: Waiting
2025-12-04T08:53:31.7102465Z 386f322edd1c: Pulling fs layer
2025-12-04T08:53:31.7102609Z 4f4fb700ef54: Pulling fs layer
2025-12-04T08:53:31.7102754Z bbe49df30697: Pulling fs layer
2025-12-04T08:53:31.7102871Z 386f322edd1c: Waiting
2025-12-04T08:53:31.7102988Z d6630aa6f375: Pulling fs layer
2025-12-04T08:53:31.7103099Z 4f4fb700ef54: Waiting
2025-12-04T08:53:31.7103203Z 6d807afc1309: Pulling fs layer
2025-12-04T08:53:31.7103318Z 60b679430e4e: Pulling fs layer
2025-12-04T08:53:31.7103428Z bbe49df30697: Waiting
2025-12-04T08:53:31.7103528Z 3992ae84f9ed: Pulling fs layer
2025-12-04T08:53:31.7103638Z 6d807afc1309: Waiting
2025-12-04T08:53:31.7103738Z d6630aa6f375: Waiting
2025-12-04T08:53:31.7103837Z 60b679430e4e: Waiting
2025-12-04T08:53:31.7103943Z 62d400609f9c: Pulling fs layer
2025-12-04T08:53:31.7104063Z 7e7b09749096: Pulling fs layer
2025-12-04T08:53:31.7104176Z 62d400609f9c: Waiting
2025-12-04T08:53:31.7104278Z 3992ae84f9ed: Waiting
2025-12-04T08:53:31.7104384Z 7dcdbd8421cb: Pulling fs layer
2025-12-04T08:53:31.7104503Z 7e7b09749096: Waiting
2025-12-04T08:53:31.7104684Z cbb12613719b: Pulling fs layer
2025-12-04T08:53:31.7104801Z e87038dce9bc: Pulling fs layer
2025-12-04T08:53:31.7104923Z e4606b636f96: Pulling fs layer
2025-12-04T08:53:31.7105048Z e87038dce9bc: Waiting
2025-12-04T08:53:31.7105155Z cbb12613719b: Waiting
2025-12-04T08:53:31.7105265Z 6f2a5d33b946: Pulling fs layer
2025-12-04T08:53:31.7105381Z e4606b636f96: Waiting
2025-12-04T08:53:31.7105487Z 6f2a5d33b946: Waiting
2025-12-04T08:53:31.7105597Z a4f2bf2f19e6: Pulling fs layer
2025-12-04T08:53:31.7105720Z 1ae00acdac56: Pulling fs layer
2025-12-04T08:53:31.7105838Z a4f2bf2f19e6: Waiting
2025-12-04T08:53:31.7105944Z 1ae00acdac56: Waiting
2025-12-04T08:53:32.2948447Z 3a5718b5258e: Download complete
2025-12-04T08:53:32.8862272Z 9d58e5257cef: Verifying Checksum
2025-12-04T08:53:32.8862590Z 9d58e5257cef: Download complete
2025-12-04T08:53:33.3847829Z 02de03a7213b: Verifying Checksum
2025-12-04T08:53:33.3848209Z 02de03a7213b: Download complete
2025-12-04T08:53:33.4690658Z fde80a645535: Verifying Checksum
2025-12-04T08:53:33.4690857Z fde80a645535: Download complete
2025-12-04T08:53:33.8915567Z 02de03a7213b: Pull complete
2025-12-04T08:53:33.8967345Z 3a5718b5258e: Pull complete
2025-12-04T08:53:33.9761894Z 6931c5f20e80: Download complete
2025-12-04T08:53:34.5707290Z dc8487f6c81c: Verifying Checksum
2025-12-04T08:53:34.5707765Z dc8487f6c81c: Download complete
2025-12-04T08:53:35.1510887Z 9748c5348f39: Verifying Checksum
2025-12-04T08:53:35.1511310Z 9748c5348f39: Download complete
2025-12-04T08:53:35.7569521Z 8539cc3f8d8a: Download complete
2025-12-04T08:53:37.3091214Z 170ea6d3edd6: Verifying Checksum
2025-12-04T08:53:37.3091671Z 170ea6d3edd6: Download complete
2025-12-04T08:53:38.0015084Z 32fbb88555c4: Verifying Checksum
2025-12-04T08:53:38.0015495Z 32fbb88555c4: Download complete
2025-12-04T08:53:40.4586295Z bf3aa2277692: Download complete
2025-12-04T08:53:41.1390202Z 80061bf5dcbb: Download complete
2025-12-04T08:53:41.7961421Z 6e9524f4518e: Verifying Checksum
2025-12-04T08:53:41.7961696Z 6e9524f4518e: Download complete
2025-12-04T08:53:42.4066225Z ce919d4bf5ee: Verifying Checksum
2025-12-04T08:53:42.4066459Z ce919d4bf5ee: Download complete
2025-12-04T08:53:44.1618517Z 3231e1ab814b: Verifying Checksum
2025-12-04T08:53:44.1619063Z 3231e1ab814b: Download complete
2025-12-04T08:53:44.6771862Z bf3aa2277692: Pull complete
2025-12-04T08:53:44.6811059Z 9d58e5257cef: Pull complete
2025-12-04T08:53:44.6862482Z fde80a645535: Pull complete
2025-12-04T08:53:44.6909279Z 6931c5f20e80: Pull complete
2025-12-04T08:53:44.7831233Z cb70fe22c9eb: Verifying Checksum
2025-12-04T08:53:44.7831492Z cb70fe22c9eb: Download complete
2025-12-04T08:53:45.4169910Z 17858e829c8c: Verifying Checksum
2025-12-04T08:53:45.4170271Z 17858e829c8c: Download complete
2025-12-04T08:53:45.8038051Z 170ea6d3edd6: Pull complete
2025-12-04T08:53:45.8078899Z dc8487f6c81c: Pull complete
2025-12-04T08:53:45.8120341Z 9748c5348f39: Pull complete
2025-12-04T08:53:45.8158993Z 8539cc3f8d8a: Pull complete
2025-12-04T08:53:46.0939436Z a63f3b4eed11: Verifying Checksum
2025-12-04T08:53:46.0939721Z a63f3b4eed11: Download complete
2025-12-04T08:53:46.7054675Z 10ab3d1afbc4: Verifying Checksum
2025-12-04T08:53:46.7054917Z 10ab3d1afbc4: Download complete
2025-12-04T08:55:02.1927324Z 47681e3e6f37: Verifying Checksum
2025-12-04T08:55:02.1927722Z 47681e3e6f37: Download complete
2025-12-04T08:55:02.8211315Z 025c90839a58: Download complete
2025-12-04T08:55:03.4464906Z 9255df5942ae: Verifying Checksum
2025-12-04T08:55:03.4466225Z 9255df5942ae: Download complete
2025-12-04T08:55:04.1049386Z f71ca9d4ed1c: Download complete
2025-12-04T08:56:31.1424298Z d02b47b56ca7: Verifying Checksum
2025-12-04T08:56:31.1427832Z d02b47b56ca7: Download complete
2025-12-04T08:56:31.8413059Z 40279492aea7: Verifying Checksum
2025-12-04T08:56:31.8413341Z 40279492aea7: Download complete
2025-12-04T08:56:32.4934428Z 33a27ce74abd: Verifying Checksum
2025-12-04T08:56:32.4934755Z 33a27ce74abd: Download complete
2025-12-04T08:56:33.1090796Z 6b66ed335d1d: Verifying Checksum
2025-12-04T08:56:33.1093172Z 6b66ed335d1d: Download complete
2025-12-04T08:56:33.7590307Z 9f010fa04118: Verifying Checksum
2025-12-04T08:56:33.7590701Z 9f010fa04118: Download complete
2025-12-04T08:56:34.9227615Z 6c64d5e8bb6a: Verifying Checksum
2025-12-04T08:56:34.9228094Z 6c64d5e8bb6a: Download complete
2025-12-04T08:56:35.4994678Z c20ea058f549: Download complete
2025-12-04T08:56:36.0933642Z 3c4fd2d54638: Verifying Checksum
2025-12-04T08:56:36.0934022Z 3c4fd2d54638: Download complete
2025-12-04T08:56:37.7610589Z 964ebac3d7a9: Verifying Checksum
2025-12-04T08:56:37.7610799Z 964ebac3d7a9: Download complete
2025-12-04T08:56:38.3593029Z 2aaa7210673f: Download complete
2025-12-04T08:56:38.9684805Z fa273daa0037: Verifying Checksum
2025-12-04T08:56:38.9685786Z fa273daa0037: Download complete
2025-12-04T08:56:40.5029721Z d931a62fd240: Verifying Checksum
2025-12-04T08:56:40.5030022Z d931a62fd240: Download complete
2025-12-04T08:56:41.1239600Z d3573d61c28e: Download complete
2025-12-04T08:56:41.7233009Z f9b32f08c490: Download complete
2025-12-04T08:56:42.3443492Z 3a0206399d60: Verifying Checksum
2025-12-04T08:56:42.3443927Z 3a0206399d60: Download complete
2025-12-04T08:56:42.9669209Z 386f322edd1c: Download complete
2025-12-04T08:56:43.2722278Z 4f4fb700ef54: Verifying Checksum
2025-12-04T08:56:43.2722651Z 4f4fb700ef54: Download complete
2025-12-04T08:56:43.8873949Z bbe49df30697: Verifying Checksum
2025-12-04T08:56:43.8874374Z bbe49df30697: Download complete
2025-12-04T08:56:44.4956504Z d6630aa6f375: Verifying Checksum
2025-12-04T08:56:44.4956769Z d6630aa6f375: Download complete
2025-12-04T08:56:45.2469283Z 6d807afc1309: Verifying Checksum
2025-12-04T08:56:45.2469730Z 6d807afc1309: Download complete
2025-12-04T08:56:45.8532851Z 60b679430e4e: Download complete
2025-12-04T08:56:46.4962796Z 3992ae84f9ed: Verifying Checksum
2025-12-04T08:56:46.4963103Z 3992ae84f9ed: Download complete
2025-12-04T08:56:47.1477057Z 62d400609f9c: Verifying Checksum
2025-12-04T08:56:47.1477504Z 62d400609f9c: Download complete
2025-12-04T08:56:47.7582435Z 7e7b09749096: Verifying Checksum
2025-12-04T08:56:47.7582863Z 7e7b09749096: Download complete
2025-12-04T09:07:24.8051552Z af88f886884f: Verifying Checksum
2025-12-04T09:07:24.8051944Z af88f886884f: Download complete
2025-12-04T09:07:25.4459741Z cbb12613719b: Download complete
2025-12-04T09:07:26.0307486Z e87038dce9bc: Verifying Checksum
2025-12-04T09:07:30.7088213Z e87038dce9bc: Download complete
2025-12-04T09:07:30.7088583Z e4606b636f96: Verifying Checksum
2025-12-04T09:07:30.7088813Z e4606b636f96: Download complete
2025-12-04T09:07:31.3182131Z 6f2a5d33b946: Verifying Checksum
2025-12-04T09:07:31.3182446Z 6f2a5d33b946: Download complete
2025-12-04T09:07:31.9003627Z a4f2bf2f19e6: Verifying Checksum
2025-12-04T09:07:31.9003888Z a4f2bf2f19e6: Download complete
2025-12-04T09:07:34.3565402Z 1ae00acdac56: Verifying Checksum
2025-12-04T09:07:34.3565853Z 1ae00acdac56: Download complete
2025-12-04T09:07:46.7096350Z af88f886884f: Pull complete
2025-12-04T09:07:46.7137950Z 32fbb88555c4: Pull complete
2025-12-04T09:07:47.2913964Z 3231e1ab814b: Pull complete
2025-12-04T09:07:47.2949896Z 80061bf5dcbb: Pull complete
2025-12-04T09:07:47.2988370Z 6e9524f4518e: Pull complete
2025-12-04T09:07:47.3035269Z ce919d4bf5ee: Pull complete
2025-12-04T09:07:50.5761110Z 47681e3e6f37: Pull complete
2025-12-04T09:07:50.5805084Z cb70fe22c9eb: Pull complete
2025-12-04T09:07:50.5864904Z 17858e829c8c: Pull complete
2025-12-04T09:07:50.5958694Z a63f3b4eed11: Pull complete
2025-12-04T09:07:50.5996913Z 10ab3d1afbc4: Pull complete
2025-12-04T09:21:14.5673826Z 7dcdbd8421cb: Verifying Checksum
2025-12-04T09:21:14.5674153Z 7dcdbd8421cb: Download complete
2025-12-04T09:32:33.9755074Z 98ca88b5095b: Verifying Checksum
2025-12-04T09:32:33.9755541Z 98ca88b5095b: Download complete
2025-12-04T09:33:18.6333926Z 98ca88b5095b: Pull complete
2025-12-04T09:33:18.6373548Z 025c90839a58: Pull complete
2025-12-04T09:33:18.6420523Z 9255df5942ae: Pull complete
2025-12-04T09:33:18.6457834Z f71ca9d4ed1c: Pull complete
2025-12-04T09:33:23.4223271Z d02b47b56ca7: Pull complete
2025-12-04T09:33:23.4262836Z 40279492aea7: Pull complete
2025-12-04T09:33:23.4305165Z 33a27ce74abd: Pull complete
2025-12-04T09:33:23.4345542Z 6b66ed335d1d: Pull complete
2025-12-04T09:33:23.4388471Z 9f010fa04118: Pull complete
2025-12-04T09:33:23.4705086Z 6c64d5e8bb6a: Pull complete
2025-12-04T09:33:23.4743507Z c20ea058f549: Pull complete
2025-12-04T09:33:23.4781201Z 3c4fd2d54638: Pull complete
2025-12-04T09:33:23.6836876Z 964ebac3d7a9: Pull complete
2025-12-04T09:33:23.6871976Z 2aaa7210673f: Pull complete
2025-12-04T09:33:23.6902203Z fa273daa0037: Pull complete
2025-12-04T09:33:23.7873194Z d931a62fd240: Pull complete
2025-12-04T09:33:23.7909274Z d3573d61c28e: Pull complete
2025-12-04T09:33:23.7996563Z f9b32f08c490: Pull complete
2025-12-04T09:33:23.8043589Z 3a0206399d60: Pull complete
2025-12-04T09:33:23.8088186Z 386f322edd1c: Pull complete
2025-12-04T09:33:23.8122923Z 4f4fb700ef54: Pull complete
2025-12-04T09:33:23.8193350Z bbe49df30697: Pull complete
2025-12-04T09:33:23.8231498Z d6630aa6f375: Pull complete
2025-12-04T09:33:23.8274196Z 6d807afc1309: Pull complete
2025-12-04T09:33:23.8308856Z 60b679430e4e: Pull complete
2025-12-04T09:33:23.8339899Z 3992ae84f9ed: Pull complete
2025-12-04T09:33:23.8431957Z 62d400609f9c: Pull complete
2025-12-04T09:33:23.8466458Z 7e7b09749096: Pull complete
2025-12-04T09:34:01.7625054Z 7dcdbd8421cb: Pull complete
2025-12-04T09:34:01.7666612Z cbb12613719b: Pull complete
2025-12-04T09:34:01.7712099Z e87038dce9bc: Pull complete
2025-12-04T09:34:04.1248773Z e4606b636f96: Pull complete
2025-12-04T09:34:04.1294732Z 6f2a5d33b946: Pull complete
2025-12-04T09:34:04.1384446Z a4f2bf2f19e6: Pull complete
2025-12-04T09:34:04.7383822Z 1ae00acdac56: Pull complete
2025-12-04T09:34:04.7406895Z Digest: sha256:f0728d30af94602d09207f794eb469a578a6cd97e72880fb3f401801d2f4acc6
2025-12-04T09:34:04.7409548Z Status: Downloaded newer image for 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T09:34:04.7416782Z 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T09:34:04.7463663Z Prepare all required actions
2025-12-04T09:34:04.7479578Z ##[group]Run ./.github/actions/get-workflow-job-id
2025-12-04T09:34:04.7479741Z with:
2025-12-04T09:34:04.7480069Z   github-token: ***
2025-12-04T09:34:04.7480176Z env:
2025-12-04T09:34:04.7480280Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:34:04.7480500Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:34:04.7480687Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:34:04.7480861Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:34:04.7481260Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:34:04.7481644Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:34:04.7481768Z   AWS_REGION: us-east-1
2025-12-04T09:34:04.7481903Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:34:04.7482089Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:34:04.7484034Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:34:04.7484155Z ##[endgroup]
2025-12-04T09:34:04.7492009Z ##[group]Run set -eux
2025-12-04T09:34:04.7492138Z set -eux
2025-12-04T09:34:04.7492316Z python3 .github/scripts/get_workflow_job_id.py "${GITHUB_RUN_ID}" "${RUNNER_NAME}"
2025-12-04T09:34:04.7496605Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T09:34:04.7496758Z env:
2025-12-04T09:34:04.7496859Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:34:04.7497001Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:34:04.7497184Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:34:04.7497356Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:34:04.7497753Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:34:04.7498125Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:34:04.7498257Z   AWS_REGION: us-east-1
2025-12-04T09:34:04.7498419Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:34:04.7498595Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:34:04.7500573Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:34:04.7500751Z   GITHUB_TOKEN: ***
2025-12-04T09:34:04.7500854Z ##[endgroup]
2025-12-04T09:34:04.7517789Z + python3 .github/scripts/get_workflow_job_id.py 19922812470 linux.rocm.gpu.gfx942.1.b-gwk9b-runner-xf6tf
2025-12-04T09:34:05.7504334Z Setting output job-id=57116139325
2025-12-04T09:34:05.7504650Z Setting output job-name=linux-noble-rocm-py3.12-mi300 / test (default, 3, 6, linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests)
2025-12-04T09:34:05.7594286Z Prepare all required actions
2025-12-04T09:34:05.7594535Z Getting action download info
2025-12-04T09:34:06.1418275Z Download action repository 'seemethere/download-artifact-s3@v4' (SHA:1da556a7aa0a088e3153970611f6c432d58e80e6)
2025-12-04T09:34:07.2522906Z Download action repository 'actions/download-artifact@v4' (SHA:d3f86a106a0bac45b974a628896c90dbdf5c8093)
2025-12-04T09:34:08.3033159Z ##[group]Run ./.github/actions/download-build-artifacts
2025-12-04T09:34:08.3033318Z with:
2025-12-04T09:34:08.3033431Z   name: linux-noble-rocm-py3.12-mi300
2025-12-04T09:34:08.3033567Z   s3-bucket: gha-artifacts
2025-12-04T09:34:08.3033679Z env:
2025-12-04T09:34:08.3033777Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:34:08.3033916Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:34:08.3034096Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:34:08.3034268Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:34:08.3034698Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:34:08.3035086Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:34:08.3035207Z   AWS_REGION: us-east-1
2025-12-04T09:34:08.3035556Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:34:08.3035774Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:34:08.3037710Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:34:08.3037822Z ##[endgroup]
2025-12-04T09:34:08.3051604Z ##[group]Run seemethere/download-artifact-s3@v4
2025-12-04T09:34:08.3051757Z with:
2025-12-04T09:34:08.3051866Z   name: linux-noble-rocm-py3.12-mi300
2025-12-04T09:34:08.3052001Z   s3-bucket: gha-artifacts
2025-12-04T09:34:08.3052116Z   region: us-east-1
2025-12-04T09:34:08.3052217Z env:
2025-12-04T09:34:08.3052314Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:34:08.3052455Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:34:08.3052642Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:34:08.3052815Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:34:08.3053200Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:34:08.3053584Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:34:08.3053705Z   AWS_REGION: us-east-1
2025-12-04T09:34:08.3053872Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:34:08.3054028Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:34:08.3055970Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:34:08.3056078Z ##[endgroup]
2025-12-04T09:34:08.5285993Z (node:17247) NOTE: We are formalizing our plans to enter AWS SDK for JavaScript (v2) into maintenance mode in 2023.
2025-12-04T09:34:08.5286223Z 
2025-12-04T09:34:08.5286314Z Please migrate your code to use AWS SDK for JavaScript (v3).
2025-12-04T09:34:08.5286596Z For more information, check the migration guide at https://a.co/7PzMCcy
2025-12-04T09:34:08.5286822Z (Use `node --trace-warnings ...` to show where the warning was created)
2025-12-04T09:34:08.8092295Z Found 1 objects with prefix pytorch/pytorch/19922812470/linux-noble-rocm-py3.12-mi300/
2025-12-04T09:34:08.8092606Z Starting download (1/1): /home/runner/_work/pytorch/pytorch/artifacts.zip
2025-12-04T09:36:51.6049177Z Finished download (1/1): /home/runner/_work/pytorch/pytorch/artifacts.zip
2025-12-04T09:36:51.6053789Z Artifact download has finished successfully
2025-12-04T09:36:51.6213183Z ##[group]Run unzip -o artifacts.zip
2025-12-04T09:36:51.6213358Z unzip -o artifacts.zip
2025-12-04T09:36:51.6217753Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T09:36:51.6217913Z env:
2025-12-04T09:36:51.6218014Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:36:51.6218157Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:36:51.6218520Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:36:51.6218694Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:36:51.6219082Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:36:51.6219467Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:36:51.6219588Z   AWS_REGION: us-east-1
2025-12-04T09:36:51.6219782Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:36:51.6219940Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:36:51.6221898Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:36:51.6222006Z ##[endgroup]
2025-12-04T09:36:51.6262478Z Archive:  artifacts.zip
2025-12-04T09:36:51.6264510Z    creating: dist/
2025-12-04T09:36:51.6346957Z   inflating: dist/.ninja_log         
2025-12-04T09:36:54.5496880Z   inflating: dist/torch-2.10.0a0+gitffd9b0f-cp312-cp312-linux_x86_64.whl  
2025-12-04T09:36:54.5497485Z    creating: build/
2025-12-04T09:36:54.5497857Z    creating: build/custom_test_artifacts/
2025-12-04T09:36:54.5498299Z    creating: build/custom_test_artifacts/custom-op-build/
2025-12-04T09:36:54.5498810Z    creating: build/custom_test_artifacts/custom-op-build/CMakeFiles/
2025-12-04T09:36:54.5499405Z    creating: build/custom_test_artifacts/custom-op-build/CMakeFiles/pkgRedirects/
2025-12-04T09:36:54.5501026Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/CMakeConfigureLog.yaml  
2025-12-04T09:36:54.5501687Z    creating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/
2025-12-04T09:36:54.5502345Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/CMakeSystem.cmake  
2025-12-04T09:36:54.5503043Z    creating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/CompilerIdC/
2025-12-04T09:36:54.5503733Z    creating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/CompilerIdC/tmp/
2025-12-04T09:36:54.5504527Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/CompilerIdC/CMakeCCompilerId.c  
2025-12-04T09:36:54.5505330Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/CompilerIdC/a.out  
2025-12-04T09:36:54.5506070Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/CMakeCCompiler.cmake  
2025-12-04T09:36:54.5506768Z    creating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/CompilerIdCXX/
2025-12-04T09:36:54.5507238Z    creating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/CompilerIdCXX/tmp/
2025-12-04T09:36:54.5507752Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/CompilerIdCXX/CMakeCXXCompilerId.cpp  
2025-12-04T09:36:54.5508276Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/CompilerIdCXX/a.out  
2025-12-04T09:36:54.5508761Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/CMakeCXXCompiler.cmake  
2025-12-04T09:36:54.5509310Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/CMakeDetermineCompilerABI_C.bin  
2025-12-04T09:36:54.5509874Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/3.31.6/CMakeDetermineCompilerABI_CXX.bin  
2025-12-04T09:36:54.5510391Z    creating: build/custom_test_artifacts/custom-op-build/CMakeFiles/CMakeScratch/
2025-12-04T09:36:54.5510832Z    creating: build/custom_test_artifacts/custom-op-build/CMakeFiles/CMakeTmp/
2025-12-04T09:36:54.5511246Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/cmake.check_cache  
2025-12-04T09:36:54.5511670Z    creating: build/custom_test_artifacts/custom-op-build/CMakeFiles/custom_ops.dir/
2025-12-04T09:36:54.5512135Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/custom_ops.dir/compiler_depend.ts  
2025-12-04T09:36:54.5512668Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/custom_ops.dir/compiler_depend.make  
2025-12-04T09:36:54.5513344Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/custom_ops.dir/depend.make  
2025-12-04T09:36:54.5513821Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/custom_ops.dir/link.txt  
2025-12-04T09:36:54.5514305Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/custom_ops.dir/cmake_clean.cmake  
2025-12-04T09:36:54.5514787Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/custom_ops.dir/build.make  
2025-12-04T09:36:54.5515281Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/custom_ops.dir/DependInfo.cmake  
2025-12-04T09:36:54.5515775Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/custom_ops.dir/flags.make  
2025-12-04T09:36:54.5516260Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/custom_ops.dir/progress.make  
2025-12-04T09:36:54.5521169Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/custom_ops.dir/op.cpp.o.d  
2025-12-04T09:36:54.5627875Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/custom_ops.dir/op.cpp.o  
2025-12-04T09:36:54.5628176Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/custom_ops.dir/link.d  
2025-12-04T09:36:54.5628528Z    creating: build/custom_test_artifacts/custom-op-build/CMakeFiles/test_custom_ops.dir/
2025-12-04T09:36:54.5628830Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/test_custom_ops.dir/compiler_depend.ts  
2025-12-04T09:36:54.5629381Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/test_custom_ops.dir/compiler_depend.make  
2025-12-04T09:36:54.5629702Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/test_custom_ops.dir/depend.make  
2025-12-04T09:36:54.5629994Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/test_custom_ops.dir/link.txt  
2025-12-04T09:36:54.5630297Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/test_custom_ops.dir/cmake_clean.cmake  
2025-12-04T09:36:54.5630656Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/test_custom_ops.dir/build.make  
2025-12-04T09:36:54.5630958Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/test_custom_ops.dir/DependInfo.cmake  
2025-12-04T09:36:54.5631254Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/test_custom_ops.dir/flags.make  
2025-12-04T09:36:54.5631549Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/test_custom_ops.dir/progress.make  
2025-12-04T09:36:54.5642792Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/test_custom_ops.dir/test_custom_ops.cpp.o.d  
2025-12-04T09:36:54.5686112Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/test_custom_ops.dir/test_custom_ops.cpp.o  
2025-12-04T09:36:54.5686485Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/test_custom_ops.dir/link.d  
2025-12-04T09:36:54.5686805Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/CMakeDirectoryInformation.cmake  
2025-12-04T09:36:54.5687110Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/TargetDirectories.txt  
2025-12-04T09:36:54.5687375Z  extracting: build/custom_test_artifacts/custom-op-build/CMakeFiles/progress.marks  
2025-12-04T09:36:54.5687626Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/Makefile2  
2025-12-04T09:36:54.5688043Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeFiles/Makefile.cmake  
2025-12-04T09:36:54.5688302Z   inflating: build/custom_test_artifacts/custom-op-build/hipblaslt_test_outer_vec.cc  
2025-12-04T09:36:54.5688546Z   inflating: build/custom_test_artifacts/custom-op-build/hipblaslt_test_vec_ext.cc  
2025-12-04T09:36:54.5689305Z   inflating: build/custom_test_artifacts/custom-op-build/CMakeCache.txt  
2025-12-04T09:36:54.5689638Z   inflating: build/custom_test_artifacts/custom-op-build/Makefile  
2025-12-04T09:36:54.5689878Z   inflating: build/custom_test_artifacts/custom-op-build/cmake_install.cmake  
2025-12-04T09:36:54.5780997Z   inflating: build/custom_test_artifacts/custom-op-build/libcustom_ops.so  
2025-12-04T09:36:54.5811045Z   inflating: build/custom_test_artifacts/custom-op-build/test_custom_ops  
2025-12-04T09:36:54.5811251Z    creating: build/custom_test_artifacts/jit-hook-build/
2025-12-04T09:36:54.5811444Z    creating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/
2025-12-04T09:36:54.5811658Z    creating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/pkgRedirects/
2025-12-04T09:36:54.5813749Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/CMakeConfigureLog.yaml  
2025-12-04T09:36:54.5813999Z    creating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/
2025-12-04T09:36:54.5814239Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/CMakeSystem.cmake  
2025-12-04T09:36:54.5814502Z    creating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/CompilerIdC/
2025-12-04T09:36:54.5814753Z    creating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/CompilerIdC/tmp/
2025-12-04T09:36:54.5815536Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/CompilerIdC/CMakeCCompilerId.c  
2025-12-04T09:36:54.5816331Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/CompilerIdC/a.out  
2025-12-04T09:36:54.5816623Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/CMakeCCompiler.cmake  
2025-12-04T09:36:54.5816947Z    creating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/CompilerIdCXX/
2025-12-04T09:36:54.5817206Z    creating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/CompilerIdCXX/tmp/
2025-12-04T09:36:54.5818148Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/CompilerIdCXX/CMakeCXXCompilerId.cpp  
2025-12-04T09:36:54.5818889Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/CompilerIdCXX/a.out  
2025-12-04T09:36:54.5819233Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/CMakeCXXCompiler.cmake  
2025-12-04T09:36:54.5820323Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/CMakeDetermineCompilerABI_C.bin  
2025-12-04T09:36:54.5821189Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/3.31.6/CMakeDetermineCompilerABI_CXX.bin  
2025-12-04T09:36:54.5821476Z    creating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/CMakeScratch/
2025-12-04T09:36:54.5821709Z    creating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/CMakeTmp/
2025-12-04T09:36:54.5821947Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/cmake.check_cache  
2025-12-04T09:36:54.5822193Z    creating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/test_jit_hooks.dir/
2025-12-04T09:36:54.5822471Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/test_jit_hooks.dir/compiler_depend.ts  
2025-12-04T09:36:54.5822789Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/test_jit_hooks.dir/compiler_depend.make  
2025-12-04T09:36:54.5823100Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/test_jit_hooks.dir/depend.make  
2025-12-04T09:36:54.5823381Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/test_jit_hooks.dir/link.txt  
2025-12-04T09:36:54.5823672Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/test_jit_hooks.dir/cmake_clean.cmake  
2025-12-04T09:36:54.5823977Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/test_jit_hooks.dir/build.make  
2025-12-04T09:36:54.5824270Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/test_jit_hooks.dir/DependInfo.cmake  
2025-12-04T09:36:54.5824565Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/test_jit_hooks.dir/flags.make  
2025-12-04T09:36:54.5824850Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/test_jit_hooks.dir/progress.make  
2025-12-04T09:36:54.5835419Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/test_jit_hooks.dir/test_jit_hooks.cpp.o.d  
2025-12-04T09:36:54.5869175Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/test_jit_hooks.dir/test_jit_hooks.cpp.o  
2025-12-04T09:36:54.5869502Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/test_jit_hooks.dir/link.d  
2025-12-04T09:36:54.5869920Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/CMakeDirectoryInformation.cmake  
2025-12-04T09:36:54.5870215Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/TargetDirectories.txt  
2025-12-04T09:36:54.5870499Z  extracting: build/custom_test_artifacts/jit-hook-build/CMakeFiles/progress.marks  
2025-12-04T09:36:54.5870743Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/Makefile2  
2025-12-04T09:36:54.5871374Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeFiles/Makefile.cmake  
2025-12-04T09:36:54.5871625Z   inflating: build/custom_test_artifacts/jit-hook-build/hipblaslt_test_outer_vec.cc  
2025-12-04T09:36:54.5871872Z   inflating: build/custom_test_artifacts/jit-hook-build/hipblaslt_test_vec_ext.cc  
2025-12-04T09:36:54.5872592Z   inflating: build/custom_test_artifacts/jit-hook-build/CMakeCache.txt  
2025-12-04T09:36:54.5872877Z   inflating: build/custom_test_artifacts/jit-hook-build/Makefile  
2025-12-04T09:36:54.5873109Z   inflating: build/custom_test_artifacts/jit-hook-build/cmake_install.cmake  
2025-12-04T09:36:54.5894053Z   inflating: build/custom_test_artifacts/jit-hook-build/test_jit_hooks  
2025-12-04T09:36:54.5894259Z    creating: build/custom_test_artifacts/custom-backend-build/
2025-12-04T09:36:54.5894462Z    creating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/
2025-12-04T09:36:54.5894701Z    creating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/pkgRedirects/
2025-12-04T09:36:54.5896641Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/CMakeConfigureLog.yaml  
2025-12-04T09:36:54.5896908Z    creating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/
2025-12-04T09:36:54.5897174Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/CMakeSystem.cmake  
2025-12-04T09:36:54.5897457Z    creating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/CompilerIdC/
2025-12-04T09:36:54.5897734Z    creating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/CompilerIdC/tmp/
2025-12-04T09:36:54.5898588Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/CompilerIdC/CMakeCCompilerId.c  
2025-12-04T09:36:54.5899480Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/CompilerIdC/a.out  
2025-12-04T09:36:54.5899784Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/CMakeCCompiler.cmake  
2025-12-04T09:36:54.5900067Z    creating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/CompilerIdCXX/
2025-12-04T09:36:54.5900355Z    creating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/CompilerIdCXX/tmp/
2025-12-04T09:36:54.5901330Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/CompilerIdCXX/CMakeCXXCompilerId.cpp  
2025-12-04T09:36:54.5901999Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/CompilerIdCXX/a.out  
2025-12-04T09:36:54.5902369Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/CMakeCXXCompiler.cmake  
2025-12-04T09:36:54.5903397Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/CMakeDetermineCompilerABI_C.bin  
2025-12-04T09:36:54.5904150Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/3.31.6/CMakeDetermineCompilerABI_CXX.bin  
2025-12-04T09:36:54.5904458Z    creating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/CMakeScratch/
2025-12-04T09:36:54.5904697Z    creating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/CMakeTmp/
2025-12-04T09:36:54.5904953Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/cmake.check_cache  
2025-12-04T09:36:54.5905291Z    creating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/custom_backend.dir/
2025-12-04T09:36:54.5905591Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/custom_backend.dir/compiler_depend.ts  
2025-12-04T09:36:54.5905925Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/custom_backend.dir/compiler_depend.make  
2025-12-04T09:36:54.5906251Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/custom_backend.dir/depend.make  
2025-12-04T09:36:54.5906551Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/custom_backend.dir/link.txt  
2025-12-04T09:36:54.5906870Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/custom_backend.dir/cmake_clean.cmake  
2025-12-04T09:36:54.5907182Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/custom_backend.dir/build.make  
2025-12-04T09:36:54.5907497Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/custom_backend.dir/DependInfo.cmake  
2025-12-04T09:36:54.5907808Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/custom_backend.dir/flags.make  
2025-12-04T09:36:54.5908116Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/custom_backend.dir/progress.make  
2025-12-04T09:36:54.5908913Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/custom_backend.dir/custom_backend.cpp.o.d  
2025-12-04T09:36:54.5972911Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/custom_backend.dir/custom_backend.cpp.o  
2025-12-04T09:36:54.5973240Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/custom_backend.dir/link.d  
2025-12-04T09:36:54.5973540Z    creating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/test_custom_backend.dir/
2025-12-04T09:36:54.5973874Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/test_custom_backend.dir/compiler_depend.ts  
2025-12-04T09:36:54.5974222Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/test_custom_backend.dir/compiler_depend.make  
2025-12-04T09:36:54.5974566Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/test_custom_backend.dir/depend.make  
2025-12-04T09:36:54.5974884Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/test_custom_backend.dir/link.txt  
2025-12-04T09:36:54.5975213Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/test_custom_backend.dir/cmake_clean.cmake  
2025-12-04T09:36:54.5975551Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/test_custom_backend.dir/build.make  
2025-12-04T09:36:54.5975881Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/test_custom_backend.dir/DependInfo.cmake  
2025-12-04T09:36:54.5976210Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/test_custom_backend.dir/flags.make  
2025-12-04T09:36:54.5976540Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/test_custom_backend.dir/progress.make  
2025-12-04T09:36:54.5987047Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/test_custom_backend.dir/test_custom_backend.cpp.o.d  
2025-12-04T09:36:54.6016502Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/test_custom_backend.dir/test_custom_backend.cpp.o  
2025-12-04T09:36:54.6016849Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/test_custom_backend.dir/link.d  
2025-12-04T09:36:54.6017177Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/CMakeDirectoryInformation.cmake  
2025-12-04T09:36:54.6017480Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/TargetDirectories.txt  
2025-12-04T09:36:54.6017760Z  extracting: build/custom_test_artifacts/custom-backend-build/CMakeFiles/progress.marks  
2025-12-04T09:36:54.6018075Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/Makefile2  
2025-12-04T09:36:54.6018443Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeFiles/Makefile.cmake  
2025-12-04T09:36:54.6018721Z   inflating: build/custom_test_artifacts/custom-backend-build/hipblaslt_test_outer_vec.cc  
2025-12-04T09:36:54.6018987Z   inflating: build/custom_test_artifacts/custom-backend-build/hipblaslt_test_vec_ext.cc  
2025-12-04T09:36:54.6019878Z   inflating: build/custom_test_artifacts/custom-backend-build/CMakeCache.txt  
2025-12-04T09:36:54.6020133Z   inflating: build/custom_test_artifacts/custom-backend-build/Makefile  
2025-12-04T09:36:54.6020618Z   inflating: build/custom_test_artifacts/custom-backend-build/cmake_install.cmake  
2025-12-04T09:36:54.6074676Z   inflating: build/custom_test_artifacts/custom-backend-build/libcustom_backend.so  
2025-12-04T09:36:54.6095597Z   inflating: build/custom_test_artifacts/custom-backend-build/test_custom_backend  
2025-12-04T09:36:54.6095854Z    creating: build/lib/
2025-12-04T09:36:54.6141063Z   inflating: build/lib/libprotobuf-lite.a  
2025-12-04T09:36:54.6386115Z   inflating: build/lib/libprotobuf.a  
2025-12-04T09:36:54.6660796Z   inflating: build/lib/libprotoc.a   
2025-12-04T09:36:54.6666094Z   inflating: build/lib/libpthreadpool.a  
2025-12-04T09:36:54.6670037Z   inflating: build/lib/libcpuinfo.a  
2025-12-04T09:36:54.6674304Z   inflating: build/lib/libcpuinfo_internals.a  
2025-12-04T09:36:54.6674710Z   inflating: build/lib/libclog.a     
2025-12-04T09:36:54.6685306Z   inflating: build/lib/libpytorch_qnnpack.a  
2025-12-04T09:36:54.6686250Z   inflating: build/lib/libnnpack_reference_layers.a  
2025-12-04T09:36:54.6787837Z   inflating: build/lib/libmicrokernels-prod.a  
2025-12-04T09:36:54.6797699Z   inflating: build/lib/libnnpack.a   
2025-12-04T09:36:54.7268461Z   inflating: build/lib/libmicrokernels-all.a  
2025-12-04T09:36:54.7306670Z   inflating: build/lib/libgtest.a    
2025-12-04T09:36:54.7316034Z   inflating: build/lib/libgmock.a    
2025-12-04T09:36:54.7316286Z   inflating: build/lib/libgtest_main.a  
2025-12-04T09:36:54.7316507Z   inflating: build/lib/libgmock_main.a  
2025-12-04T09:36:54.7366095Z   inflating: build/lib/libXNNPACK.a  
2025-12-04T09:36:54.7408660Z   inflating: build/lib/libbenchmark.a  
2025-12-04T09:36:54.7409083Z   inflating: build/lib/libbenchmark_main.a  
2025-12-04T09:36:54.7444526Z   inflating: build/lib/libasmjit.a   
2025-12-04T09:36:54.7444908Z   inflating: build/lib/libjitprofiling.a  
2025-12-04T09:36:54.8073156Z   inflating: build/lib/libfbgemm.a   
2025-12-04T09:36:54.8076810Z   inflating: build/lib/libittnotify.a  
2025-12-04T09:36:54.8093499Z   inflating: build/lib/libtensorpipe_uv.a  
2025-12-04T09:36:54.8390009Z   inflating: build/lib/libtensorpipe.a  
2025-12-04T09:36:54.8455953Z   inflating: build/lib/libgloo.a     
2025-12-04T09:36:54.8481682Z   inflating: build/lib/libonnx_proto.a  
2025-12-04T09:36:54.8702666Z   inflating: build/lib/libgloo_hip.a  
2025-12-04T09:36:54.9096079Z   inflating: build/lib/libonnx.a     
2025-12-04T09:36:54.9106900Z   inflating: build/lib/libfmt.a      
2025-12-04T09:36:55.4722368Z   inflating: build/lib/libdnnl.a     
2025-12-04T09:36:55.4892677Z   inflating: build/lib/libkineto.a   
2025-12-04T09:36:55.4957675Z   inflating: build/lib/libc10.so     
2025-12-04T09:36:55.4958905Z   inflating: build/lib/libcaffe2_nvrtc.so  
2025-12-04T09:36:55.4959308Z   inflating: build/lib/libtorch_global_deps.so  
2025-12-04T09:36:55.4984729Z   inflating: build/lib/libc10_hip.so  
2025-12-04T09:36:55.5252261Z   inflating: build/lib/libfbgemm_genai.a  
2025-12-04T09:36:57.2248938Z   inflating: build/lib/libtorch_cpu.so  
2025-12-04T09:36:57.2250614Z   inflating: build/lib/libshm.so     
2025-12-04T09:36:58.0514342Z   inflating: build/lib/libtorch_hip.so  
2025-12-04T09:36:58.0514569Z   inflating: build/lib/libtorch.so   
2025-12-04T09:36:58.0525785Z   inflating: build/lib/libjitbackend_test.so  
2025-12-04T09:36:58.0565067Z   inflating: build/lib/libtorchbind_test.so  
2025-12-04T09:36:58.0578912Z   inflating: build/lib/libbackend_with_compiler.so  
2025-12-04T09:36:58.0592962Z   inflating: build/lib/libaoti_custom_ops.so  
2025-12-04T09:36:58.1903013Z   inflating: build/lib/libtorch_python.so  
2025-12-04T09:36:58.1922971Z   inflating: build/lib/libnnapi_backend.so  
2025-12-04T09:36:58.1923141Z    creating: build/bin/
2025-12-04T09:36:58.1923328Z    creating: build/bin/CMakeFiles/
2025-12-04T09:36:58.1923581Z   inflating: build/bin/cmake_install.cmake  
2025-12-04T09:36:58.1923781Z   inflating: build/bin/CTestTestfile.cmake  
2025-12-04T09:36:58.2177037Z   inflating: build/bin/protoc-3.13.0.0  
2025-12-04T09:36:58.2429243Z   inflating: build/bin/protoc        
2025-12-04T09:36:58.2462024Z   inflating: build/bin/c10_AllocatorConfig_test  
2025-12-04T09:36:58.2492777Z   inflating: build/bin/c10_CompileTimeFunctionPointer_test  
2025-12-04T09:36:58.2524235Z   inflating: build/bin/c10_DeviceGuard_test  
2025-12-04T09:36:58.2555912Z   inflating: build/bin/c10_Device_test  
2025-12-04T09:36:58.2592092Z   inflating: build/bin/c10_DispatchKeySet_test  
2025-12-04T09:36:58.2624804Z   inflating: build/bin/c10_Scalar_test  
2025-12-04T09:36:58.2654858Z   inflating: build/bin/c10_StreamGuard_test  
2025-12-04T09:36:58.2689312Z   inflating: build/bin/c10_SymInt_test  
2025-12-04T09:36:58.2722453Z   inflating: build/bin/c10_InlineDeviceGuard_test  
2025-12-04T09:36:58.2756604Z   inflating: build/bin/c10_SizesAndStrides_test  
2025-12-04T09:36:58.2786719Z   inflating: build/bin/c10_ConstexprCrc_test  
2025-12-04T09:36:58.2828637Z   inflating: build/bin/c10_cow_test  
2025-12-04T09:36:58.2862597Z   inflating: build/bin/c10_InlineStreamGuard_test  
2025-12-04T09:36:58.2892867Z   inflating: build/bin/c10_ArrayRef_test  
2025-12-04T09:36:58.2925154Z   inflating: build/bin/c10_Bitset_test  
2025-12-04T09:36:58.2955613Z   inflating: build/bin/c10_DeadlockDetection_test  
2025-12-04T09:36:58.2990115Z   inflating: build/bin/c10_Enumerate_test  
2025-12-04T09:36:58.3021322Z   inflating: build/bin/c10_Half_test  
2025-12-04T09:36:58.3055396Z   inflating: build/bin/c10_LeftRight_test  
2025-12-04T09:36:58.3087935Z   inflating: build/bin/c10_NetworkFlow_test  
2025-12-04T09:36:58.3120301Z   inflating: build/bin/c10_IntrusiveList_test  
2025-12-04T09:36:58.3150986Z   inflating: build/bin/c10_Synchronized_test  
2025-12-04T09:36:58.3181362Z   inflating: build/bin/c10_Semaphore_test  
2025-12-04T09:36:58.3213093Z   inflating: build/bin/c10_TypeIndex_test  
2025-12-04T09:36:58.3246762Z   inflating: build/bin/c10_ThreadLocal_test  
2025-12-04T09:36:58.3278396Z   inflating: build/bin/c10_accumulate_test  
2025-12-04T09:36:58.3312213Z   inflating: build/bin/c10_bfloat16_test  
2025-12-04T09:36:58.3342952Z   inflating: build/bin/c10_bit_cast_test  
2025-12-04T09:36:58.3377727Z   inflating: build/bin/c10_complex_math_test  
2025-12-04T09:36:58.3409918Z   inflating: build/bin/c10_exception_test  
2025-12-04T09:36:58.3440286Z   inflating: build/bin/c10_error_test  
2025-12-04T09:36:58.3473813Z   inflating: build/bin/c10_complex_test  
2025-12-04T09:36:58.3504847Z   inflating: build/bin/c10_flags_test  
2025-12-04T09:36:58.3535783Z   inflating: build/bin/c10_generic_math_test  
2025-12-04T09:36:58.3570687Z   inflating: build/bin/c10_logging_test  
2025-12-04T09:36:58.3601298Z   inflating: build/bin/c10_nofatal_test  
2025-12-04T09:36:58.3632533Z   inflating: build/bin/c10_irange_test  
2025-12-04T09:36:58.3721992Z   inflating: build/bin/c10_intrusive_ptr_test  
2025-12-04T09:36:58.3754590Z   inflating: build/bin/c10_lazy_test  
2025-12-04T09:36:58.3799195Z   inflating: build/bin/c10_optional_test  
2025-12-04T09:36:58.3832061Z   inflating: build/bin/c10_registry_test  
2025-12-04T09:36:58.3866230Z   inflating: build/bin/c10_string_util_test  
2025-12-04T09:36:58.3897720Z   inflating: build/bin/c10_ssize_test  
2025-12-04T09:36:58.3986404Z   inflating: build/bin/c10_small_vector_test  
2025-12-04T09:36:58.4023949Z   inflating: build/bin/c10_ordered_preserving_dict_test  
2025-12-04T09:36:58.4054021Z   inflating: build/bin/c10_string_view_test  
2025-12-04T09:36:58.4084918Z   inflating: build/bin/c10_tempfile_test  
2025-12-04T09:36:58.4111538Z   inflating: build/bin/c10_intrusive_ptr_benchmark  
2025-12-04T09:36:58.4145507Z   inflating: build/bin/c10_typeid_test  
2025-12-04T09:36:58.4175818Z   inflating: build/bin/c10_hip_HIPAssertionsTest_1_var_test  
2025-12-04T09:36:58.4205926Z   inflating: build/bin/c10_hip_HIPAssertionsTest_catches_stream  
2025-12-04T09:36:58.4236073Z   inflating: build/bin/c10_hip_HIPAssertionsTest_catches_thread_and_block_and_device  
2025-12-04T09:36:58.4266139Z   inflating: build/bin/c10_hip_HIPAssertionsTest_from_2_processes  
2025-12-04T09:36:58.4296255Z   inflating: build/bin/c10_hip_HIPAssertionsTest_multiple_writes_from_blocks_and_threads  
2025-12-04T09:36:58.4326224Z   inflating: build/bin/c10_hip_HIPAssertionsTest_multiple_writes_from_multiple_blocks  
2025-12-04T09:36:58.4356286Z   inflating: build/bin/c10_hip_HIPAssertionsTest_multiple_writes_from_same_block  
2025-12-04T09:36:58.4386479Z   inflating: build/bin/c10_hip_HIPTest  
2025-12-04T09:36:58.4713992Z   inflating: build/bin/vec_test_all_types_DEFAULT  
2025-12-04T09:36:58.5048382Z   inflating: build/bin/vec_test_all_types_AVX512  
2025-12-04T09:36:58.5391185Z   inflating: build/bin/vec_test_all_types_AVX2  
2025-12-04T09:36:58.5448112Z   inflating: build/bin/test_aoti_abi_check  
2025-12-04T09:36:58.5478460Z   inflating: build/bin/test_vec_half_DEFAULT  
2025-12-04T09:36:58.5509094Z   inflating: build/bin/test_vec_half_AVX512  
2025-12-04T09:36:58.5539829Z   inflating: build/bin/test_vec_half_AVX2  
2025-12-04T09:36:58.5571930Z   inflating: build/bin/BackoffTest   
2025-12-04T09:36:58.5604431Z   inflating: build/bin/FileStoreTest  
2025-12-04T09:36:58.5639083Z   inflating: build/bin/TCPStoreTest  
2025-12-04T09:36:58.5671909Z   inflating: build/bin/HashStoreTest  
2025-12-04T09:36:58.5712192Z   inflating: build/bin/ProcessGroupGlooTest  
2025-12-04T09:36:58.5713678Z   inflating: build/bin/example_allreduce  
2025-12-04T09:36:58.5715656Z   inflating: build/bin/torch_shm_manager  
2025-12-04T09:36:58.5748794Z   inflating: build/bin/static_runtime_bench  
2025-12-04T09:36:58.5891947Z   inflating: build/bin/static_runtime_test  
2025-12-04T09:36:58.5935265Z   inflating: build/bin/Dict_test     
2025-12-04T09:36:58.5967205Z   inflating: build/bin/Dimname_test  
2025-12-04T09:36:58.6006293Z   inflating: build/bin/MaybeOwned_test  
2025-12-04T09:36:58.6040948Z   inflating: build/bin/NamedTensor_test  
2025-12-04T09:36:58.6076923Z   inflating: build/bin/apply_utils_test  
2025-12-04T09:36:58.6112615Z   inflating: build/bin/atest         
2025-12-04T09:36:58.6151251Z   inflating: build/bin/basic         
2025-12-04T09:36:58.6184318Z   inflating: build/bin/broadcast_test  
2025-12-04T09:36:58.6215563Z   inflating: build/bin/cpu_allocator_test  
2025-12-04T09:36:58.6250902Z   inflating: build/bin/cpu_generator_test  
2025-12-04T09:36:58.6283103Z   inflating: build/bin/cpu_profiling_allocator_test  
2025-12-04T09:36:58.6338047Z   inflating: build/bin/cpu_rng_test  
2025-12-04T09:36:58.6369973Z   inflating: build/bin/dlconvertor_test  
2025-12-04T09:36:58.6404981Z   inflating: build/bin/extension_backend_test  
2025-12-04T09:36:58.6438812Z   inflating: build/bin/half_test     
2025-12-04T09:36:58.6496548Z   inflating: build/bin/ivalue_test   
2025-12-04T09:36:58.6527019Z   inflating: build/bin/lazy_tensor_test  
2025-12-04T09:36:58.6559382Z   inflating: build/bin/math_kernel_test  
2025-12-04T09:36:58.6591645Z   inflating: build/bin/memory_format_test  
2025-12-04T09:36:58.6624313Z   inflating: build/bin/memory_overlapping_test  
2025-12-04T09:36:58.6656780Z   inflating: build/bin/mobile_memory_cleanup  
2025-12-04T09:36:58.6690751Z   inflating: build/bin/native_test   
2025-12-04T09:36:58.6722072Z   inflating: build/bin/operator_name_test  
2025-12-04T09:36:58.6753200Z   inflating: build/bin/operators_test  
2025-12-04T09:36:58.6785148Z   inflating: build/bin/packedtensoraccessor_test  
2025-12-04T09:36:58.6825717Z   inflating: build/bin/pow_test      
2025-12-04T09:36:58.6860071Z   inflating: build/bin/quantized_test  
2025-12-04T09:36:58.6890977Z   inflating: build/bin/reduce_ops_test  
2025-12-04T09:36:58.6922125Z   inflating: build/bin/reportMemoryUsage_test  
2025-12-04T09:36:58.6956085Z   inflating: build/bin/scalar_tensor_test  
2025-12-04T09:36:58.6987880Z   inflating: build/bin/stride_properties_test  
2025-12-04T09:36:58.7022830Z   inflating: build/bin/scalar_test   
2025-12-04T09:36:58.7054423Z   inflating: build/bin/StorageUtils_test  
2025-12-04T09:36:58.7087712Z   inflating: build/bin/type_ptr_test  
2025-12-04T09:36:58.7118571Z   inflating: build/bin/thread_init_test  
2025-12-04T09:36:58.7166653Z   inflating: build/bin/tensor_iterator_test  
2025-12-04T09:36:58.7199612Z   inflating: build/bin/test_parallel  
2025-12-04T09:36:58.7235603Z   inflating: build/bin/type_test     
2025-12-04T09:36:58.7267589Z   inflating: build/bin/undefined_tensor_test  
2025-12-04T09:36:58.7297965Z   inflating: build/bin/verify_api_visibility  
2025-12-04T09:36:58.7329390Z   inflating: build/bin/weakref_test  
2025-12-04T09:36:58.7371881Z   inflating: build/bin/legacy_vmap_test  
2025-12-04T09:36:58.7403387Z   inflating: build/bin/wrapdim_test  
2025-12-04T09:36:58.7434936Z   inflating: build/bin/xla_tensor_test  
2025-12-04T09:36:58.7470901Z   inflating: build/bin/IListRef_test  
2025-12-04T09:36:58.7532588Z   inflating: build/bin/List_test     
2025-12-04T09:36:58.7602485Z   inflating: build/bin/kernel_function_legacy_test  
2025-12-04T09:36:58.7642511Z   inflating: build/bin/KernelFunction_test  
2025-12-04T09:36:58.7699096Z   inflating: build/bin/kernel_function_test  
2025-12-04T09:36:58.7759103Z   inflating: build/bin/kernel_lambda_test  
2025-12-04T09:36:58.7832915Z   inflating: build/bin/kernel_lambda_legacy_test  
2025-12-04T09:36:58.7869514Z   inflating: build/bin/kernel_stackbased_test  
2025-12-04T09:36:58.7901252Z   inflating: build/bin/CppSignature_test  
2025-12-04T09:36:58.7958303Z   inflating: build/bin/make_boxed_from_unboxed_functor_test  
2025-12-04T09:36:58.7988850Z   inflating: build/bin/op_allowlist_test  
2025-12-04T09:36:58.8166392Z   inflating: build/bin/op_registration_test  
2025-12-04T09:36:58.8200056Z   inflating: build/bin/backend_fallback_test  
2025-12-04T09:36:58.8230381Z   inflating: build/bin/hip_complex_math_test  
2025-12-04T09:36:58.8271191Z   inflating: build/bin/inline_container_test  
2025-12-04T09:36:58.8301226Z   inflating: build/bin/hip_complex_test  
2025-12-04T09:36:58.8333587Z   inflating: build/bin/hip_apply_test  
2025-12-04T09:36:58.8363904Z   inflating: build/bin/hip_distributions_test  
2025-12-04T09:36:58.8394006Z   inflating: build/bin/hip_generator_test  
2025-12-04T09:36:58.8424102Z   inflating: build/bin/hip_half_test  
2025-12-04T09:36:58.8454190Z   inflating: build/bin/hip_integer_divider_test  
2025-12-04T09:36:58.8484172Z   inflating: build/bin/hip_optional_test  
2025-12-04T09:36:58.8514284Z   inflating: build/bin/hip_packedtensoraccessor_test  
2025-12-04T09:36:58.8545974Z   inflating: build/bin/hip_dlconvertor_test  
2025-12-04T09:36:58.8576071Z   inflating: build/bin/hip_vectorized_test  
2025-12-04T09:36:58.9196008Z   inflating: build/bin/test_jit      
2025-12-04T09:36:58.9393943Z   inflating: build/bin/test_lazy     
2025-12-04T09:36:58.9427652Z   inflating: build/bin/test_dist_autograd  
2025-12-04T09:36:58.9468854Z   inflating: build/bin/test_cpp_rpc  
2025-12-04T09:36:59.0125463Z   inflating: build/bin/test_api      
2025-12-04T09:36:59.0126164Z   inflating: build/bin/parallel_benchmark  
2025-12-04T09:36:59.0126566Z    creating: .additional_ci_files/
2025-12-04T09:36:59.0162671Z   inflating: .additional_ci_files/test-times.json  
2025-12-04T09:36:59.0294602Z   inflating: .additional_ci_files/test-class-times.json  
2025-12-04T09:36:59.0319357Z ##[group]Run rm artifacts.zip
2025-12-04T09:36:59.0319513Z rm artifacts.zip
2025-12-04T09:36:59.0328437Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T09:36:59.0328591Z env:
2025-12-04T09:36:59.0328695Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:36:59.0328995Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:36:59.0329178Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:36:59.0329348Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:36:59.0329734Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:36:59.0330116Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:36:59.0330235Z   AWS_REGION: us-east-1
2025-12-04T09:36:59.0330403Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:36:59.0330595Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:36:59.0332530Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:36:59.0332639Z ##[endgroup]
2025-12-04T09:36:59.1143490Z ##[group]Run df -H
2025-12-04T09:36:59.1143597Z df -H
2025-12-04T09:36:59.1146130Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T09:36:59.1146276Z env:
2025-12-04T09:36:59.1146377Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:36:59.1146513Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:36:59.1146688Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:36:59.1146853Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:36:59.1147231Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:36:59.1147730Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:36:59.1147845Z   AWS_REGION: us-east-1
2025-12-04T09:36:59.1147987Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:36:59.1148145Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:36:59.1150032Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:36:59.1150135Z ##[endgroup]
2025-12-04T09:36:59.1467992Z Filesystem      Size  Used Avail Use% Mounted on
2025-12-04T09:36:59.1468350Z overlay          16T  778G   15T   6% /
2025-12-04T09:36:59.1468745Z tmpfs            68M     0   68M   0% /dev
2025-12-04T09:36:59.1469200Z /dev/md0         16T  778G   15T   6% /run
2025-12-04T09:36:59.1469469Z shm              68M  4.1k   68M   1% /dev/shm
2025-12-04T09:36:59.1469800Z amdprj2-k8s_2   5.5T  120G  5.4T   3% /home/runner/pytorch-data
2025-12-04T09:36:59.1470190Z tmpfs           3.3T   13k  3.3T   1% /run/secrets/kubernetes.io/serviceaccount
2025-12-04T09:36:59.1470729Z tmpfs           1.7T     0  1.7T   0% /proc/acpi
2025-12-04T09:36:59.1471009Z tmpfs           1.7T     0  1.7T   0% /proc/scsi
2025-12-04T09:36:59.1471282Z tmpfs           1.7T     0  1.7T   0% /sys/firmware
2025-12-04T09:36:59.1471597Z tmpfs           1.7T     0  1.7T   0% /sys/devices/virtual/powercap
2025-12-04T09:36:59.1499547Z Prepare all required actions
2025-12-04T09:36:59.1499789Z Getting action download info
2025-12-04T09:36:59.5572661Z ##[group]Run ./.github/actions/download-td-artifacts
2025-12-04T09:36:59.5572820Z with:
2025-12-04T09:36:59.5572915Z env:
2025-12-04T09:36:59.5573010Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:36:59.5573151Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:36:59.5573329Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:36:59.5573495Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:36:59.5573883Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:36:59.5574267Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:36:59.5574382Z   AWS_REGION: us-east-1
2025-12-04T09:36:59.5574578Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:36:59.5574777Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:36:59.5576713Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:36:59.5576817Z ##[endgroup]
2025-12-04T09:36:59.5589756Z ##[group]Run seemethere/download-artifact-s3@v4
2025-12-04T09:36:59.5589892Z with:
2025-12-04T09:36:59.5589984Z   name: td_results
2025-12-04T09:36:59.5590085Z   s3-bucket: gha-artifacts
2025-12-04T09:36:59.5590193Z   region: us-east-1
2025-12-04T09:36:59.5590287Z env:
2025-12-04T09:36:59.5590380Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:36:59.5590554Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:36:59.5590730Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:36:59.5590898Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:36:59.5591282Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:36:59.5591650Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:36:59.5591768Z   AWS_REGION: us-east-1
2025-12-04T09:36:59.5591900Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:36:59.5592053Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:36:59.5593964Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:36:59.5594066Z ##[endgroup]
2025-12-04T09:36:59.7857115Z (node:17286) NOTE: We are formalizing our plans to enter AWS SDK for JavaScript (v2) into maintenance mode in 2023.
2025-12-04T09:36:59.7857382Z 
2025-12-04T09:36:59.7857493Z Please migrate your code to use AWS SDK for JavaScript (v3).
2025-12-04T09:36:59.7857777Z For more information, check the migration guide at https://a.co/7PzMCcy
2025-12-04T09:36:59.7858343Z (Use `node --trace-warnings ...` to show where the warning was created)
2025-12-04T09:37:00.0592036Z Found 1 objects with prefix pytorch/pytorch/19922812470/td_results/
2025-12-04T09:37:00.0592448Z Starting download (1/1): /home/runner/_work/pytorch/pytorch/td_results.json
2025-12-04T09:37:00.3891414Z Finished download (1/1): /home/runner/_work/pytorch/pytorch/td_results.json
2025-12-04T09:37:00.3895926Z Artifact download has finished successfully
2025-12-04T09:37:00.4059419Z ##[group]Run mkdir -p .additional_ci_files
2025-12-04T09:37:00.4059627Z mkdir -p .additional_ci_files
2025-12-04T09:37:00.4059846Z mv td_results.json .additional_ci_files/td_results.json || true
2025-12-04T09:37:00.4064849Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T09:37:00.4065037Z env:
2025-12-04T09:37:00.4065160Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:37:00.4065336Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:37:00.4065586Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:37:00.4065799Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:37:00.4066489Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:37:00.4066873Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:37:00.4066992Z   AWS_REGION: us-east-1
2025-12-04T09:37:00.4067181Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:37:00.4067341Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:37:00.4069332Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:37:00.4069440Z ##[endgroup]
2025-12-04T09:37:00.4138055Z ##[group]Run .github/scripts/parse_ref.py
2025-12-04T09:37:00.4138247Z .github/scripts/parse_ref.py
2025-12-04T09:37:00.4144957Z shell: /usr/bin/bash -e {0}
2025-12-04T09:37:00.4145085Z env:
2025-12-04T09:37:00.4145195Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:37:00.4145375Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:37:00.4145585Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:37:00.4145785Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:37:00.4146236Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:37:00.4146605Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:37:00.4146723Z   AWS_REGION: us-east-1
2025-12-04T09:37:00.4146884Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:37:00.4147057Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:37:00.4148994Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:37:00.4149098Z ##[endgroup]
2025-12-04T09:37:00.4254675Z Setting output branch=main
2025-12-04T09:37:00.4326135Z Prepare all required actions
2025-12-04T09:37:00.4326357Z Getting action download info
2025-12-04T09:37:00.6502081Z ##[group]Run ./.github/actions/filter-test-configs
2025-12-04T09:37:00.6502235Z with:
2025-12-04T09:37:00.6502467Z   github-token: ***
2025-12-04T09:37:00.6504147Z   test-matrix: {"include": [{"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}]}
2025-12-04T09:37:00.6506123Z   job-name: linux-noble-rocm-py3.12-mi300 / test (default, 3, 6, linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests)
2025-12-04T09:37:00.6506335Z env:
2025-12-04T09:37:00.6506427Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:37:00.6506570Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:37:00.6506747Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:37:00.6506911Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:37:00.6507298Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:37:00.6507663Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:37:00.6507778Z   AWS_REGION: us-east-1
2025-12-04T09:37:00.6507901Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:37:00.6508050Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:37:00.6509947Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:37:00.6510072Z ##[endgroup]
2025-12-04T09:37:00.6525822Z ##[group]Run nick-fields/retry@v3.0.0
2025-12-04T09:37:00.6525952Z with:
2025-12-04T09:37:00.6526045Z   shell: bash
2025-12-04T09:37:00.6526151Z   timeout_minutes: 10
2025-12-04T09:37:00.6526257Z   max_attempts: 5
2025-12-04T09:37:00.6526361Z   retry_wait_seconds: 30
2025-12-04T09:37:00.6526667Z   command: set -eux
# PyYAML 6.0 doesn't work with MacOS x86 anymore
# This must run on Python-3.7 (AmazonLinux2) so can't use request=3.32.2
python3 -m pip install requests==2.27.1 pyyaml==6.0.2

2025-12-04T09:37:00.6526972Z   polling_interval_seconds: 1
2025-12-04T09:37:00.6527108Z   warning_on_retry: true
2025-12-04T09:37:00.6527219Z   continue_on_error: false
2025-12-04T09:37:00.6527362Z env:
2025-12-04T09:37:00.6527458Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:37:00.6527612Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:37:00.6527793Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:37:00.6527965Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:37:00.6528492Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:37:00.6528862Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:37:00.6528983Z   AWS_REGION: us-east-1
2025-12-04T09:37:00.6529120Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:37:00.6529274Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:37:00.6531234Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:37:00.6531404Z   GITHUB_TOKEN: ***
2025-12-04T09:37:00.6531507Z ##[endgroup]
2025-12-04T09:37:00.6929132Z + python3 -m pip install requests==2.27.1 pyyaml==6.0.2
2025-12-04T09:37:00.8349504Z Defaulting to user installation because normal site-packages is not writeable
2025-12-04T09:37:00.9270068Z Collecting requests==2.27.1
2025-12-04T09:37:00.9592733Z   Downloading requests-2.27.1-py2.py3-none-any.whl (63 kB)
2025-12-04T09:37:00.9693187Z      ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 63.1/63.1 KB 6.2 MB/s eta 0:00:00
2025-12-04T09:37:01.0150922Z Collecting pyyaml==6.0.2
2025-12-04T09:37:01.0214571Z   Downloading PyYAML-6.0.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (751 kB)
2025-12-04T09:37:01.0699077Z      ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 751.2/751.2 KB 16.0 MB/s eta 0:00:00
2025-12-04T09:37:01.1683702Z Collecting charset-normalizer~=2.0.0
2025-12-04T09:37:01.1738368Z   Downloading charset_normalizer-2.0.12-py3-none-any.whl (39 kB)
2025-12-04T09:37:01.2068931Z Collecting urllib3<1.27,>=1.21.1
2025-12-04T09:37:01.2118829Z   Downloading urllib3-1.26.20-py2.py3-none-any.whl (144 kB)
2025-12-04T09:37:01.2205250Z      ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 144.2/144.2 KB 18.3 MB/s eta 0:00:00
2025-12-04T09:37:01.2408787Z Collecting certifi>=2017.4.17
2025-12-04T09:37:01.2459322Z   Downloading certifi-2025.11.12-py3-none-any.whl (159 kB)
2025-12-04T09:37:01.2540826Z      ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 159.4/159.4 KB 21.1 MB/s eta 0:00:00
2025-12-04T09:37:01.2683600Z Collecting idna<4,>=2.5
2025-12-04T09:37:01.2734799Z   Downloading idna-3.11-py3-none-any.whl (71 kB)
2025-12-04T09:37:01.2759403Z      ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 71.0/71.0 KB 44.3 MB/s eta 0:00:00
2025-12-04T09:37:01.3360386Z Installing collected packages: urllib3, pyyaml, idna, charset-normalizer, certifi, requests
2025-12-04T09:37:01.4498153Z   WARNING: The script normalizer is installed in '/home/runner/.local/bin' which is not on PATH.
2025-12-04T09:37:01.4499031Z   Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
2025-12-04T09:37:01.4828573Z Successfully installed certifi-2025.11.12 charset-normalizer-2.0.12 idna-3.11 pyyaml-6.0.2 requests-2.27.1 urllib3-1.26.20
2025-12-04T09:37:01.6933262Z Command completed after 1 attempt(s).
2025-12-04T09:37:01.6978614Z ##[group]Run set -x
2025-12-04T09:37:01.6978757Z set -x
2025-12-04T09:37:01.6978864Z 
2025-12-04T09:37:01.6979059Z # Use relative path here as this could be checked out anywhere, not necessarily
2025-12-04T09:37:01.6979261Z # in runner workspace
2025-12-04T09:37:01.6979453Z python3 "${GITHUB_ACTION_PATH}/../../scripts/parse_ref.py"
2025-12-04T09:37:01.6983709Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T09:37:01.6983876Z env:
2025-12-04T09:37:01.6983986Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:37:01.6984134Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:37:01.6984326Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:37:01.6984509Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:37:01.6984929Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:37:01.6985333Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:37:01.6985460Z   AWS_REGION: us-east-1
2025-12-04T09:37:01.6985767Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:37:01.6985937Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:37:01.6987848Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:37:01.6987961Z ##[endgroup]
2025-12-04T09:37:01.7006559Z + python3 /home/runner/_work/pytorch/pytorch/./.github/actions/filter-test-configs/../../scripts/parse_ref.py
2025-12-04T09:37:01.7090523Z Setting output branch=main
2025-12-04T09:37:01.7125455Z ##[group]Run echo "Workflow: ${GITHUB_WORKFLOW}"
2025-12-04T09:37:01.7125662Z echo "Workflow: ${GITHUB_WORKFLOW}"
2025-12-04T09:37:01.7125821Z echo "Job name: ${JOB_NAME}"
2025-12-04T09:37:01.7125958Z 
2025-12-04T09:37:01.7126143Z # Use relative path here as this could be checked out anywhere, not necessarily
2025-12-04T09:37:01.7126326Z # in runner workspace
2025-12-04T09:37:01.7126504Z python3 "${GITHUB_ACTION_PATH}/../../scripts/filter_test_configs.py" \
2025-12-04T09:37:01.7126699Z   --workflow "${GITHUB_WORKFLOW}" \
2025-12-04T09:37:01.7126845Z   --job-name "${JOB_NAME}" \
2025-12-04T09:37:01.7128530Z   --test-matrix "{"include": [{"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}]}" \
2025-12-04T09:37:01.7130516Z   --selected-test-configs "" \
2025-12-04T09:37:01.7130647Z   --pr-number "${PR_NUMBER}" \
2025-12-04T09:37:01.7130771Z   --tag "${TAG}" \
2025-12-04T09:37:01.7130887Z   --event-name "${EVENT_NAME}" \
2025-12-04T09:37:01.7131015Z   --schedule "${SCHEDULE}" \
2025-12-04T09:37:01.7131147Z   --branch "${HEAD_BRANCH}"
2025-12-04T09:37:01.7135521Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T09:37:01.7135673Z env:
2025-12-04T09:37:01.7135768Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:37:01.7135904Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:37:01.7136078Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:37:01.7136247Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:37:01.7136628Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:37:01.7136995Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:37:01.7137109Z   AWS_REGION: us-east-1
2025-12-04T09:37:01.7137294Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:37:01.7137445Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:37:01.7139469Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:37:01.7139695Z   GITHUB_TOKEN: ***
2025-12-04T09:37:01.7139898Z   JOB_NAME: linux-noble-rocm-py3.12-mi300 / test (default, 3, 6, linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests)
2025-12-04T09:37:01.7140108Z   PR_NUMBER: 
2025-12-04T09:37:01.7140198Z   TAG: 
2025-12-04T09:37:01.7140284Z   EVENT_NAME: schedule
2025-12-04T09:37:01.7140385Z   SCHEDULE: 29 8 * * *
2025-12-04T09:37:01.7140663Z   HEAD_BRANCH: main
2025-12-04T09:37:01.7140759Z ##[endgroup]
2025-12-04T09:37:01.7160897Z Workflow: rocm-mi300
2025-12-04T09:37:01.7161270Z Job name: linux-noble-rocm-py3.12-mi300 / test (default, 3, 6, linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests)
2025-12-04T09:37:02.3221566Z Setting output keep-going=True
2025-12-04T09:37:02.3221969Z Setting output ci-verbose-test-logs=False
2025-12-04T09:37:02.3222274Z Setting output ci-test-showlocals=False
2025-12-04T09:37:02.3222561Z Setting output ci-no-test-timeout=False
2025-12-04T09:37:02.3222858Z Setting output ci-no-td=False
2025-12-04T09:37:02.3223122Z Setting output ci-td-distributed=False
2025-12-04T09:37:02.3223975Z Setting output is-unstable=False
2025-12-04T09:37:02.3224236Z Setting output reenabled-issues=
2025-12-04T09:37:02.3232178Z Setting output test-matrix={"include": [{"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}]}
2025-12-04T09:37:02.3238980Z Setting output is-test-matrix-empty=False
2025-12-04T09:37:02.3338676Z ##[group]Run echo "Filtered matrix:"
2025-12-04T09:37:02.3338863Z echo "Filtered matrix:"
2025-12-04T09:37:02.3343992Z echo "{"include": [{"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 1, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 2, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 3, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 4, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 5, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "mem_leak_check": "mem_leak_check", "rerun_disabled_tests": "rerun_disabled_tests"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests", "mem_leak_check": "mem_leak_check"}, {"config": "default", "shard": 6, "num_shards": 6, "runner": "linux.rocm.gpu.gfx942.1.b", "rerun_disabled_tests": "rerun_disabled_tests"}]}"
2025-12-04T09:37:02.3348081Z 
2025-12-04T09:37:02.3348167Z echo
2025-12-04T09:37:02.3348279Z echo "Is the current job unstable? False"
2025-12-04T09:37:02.3348409Z 
2025-12-04T09:37:02.3348489Z echo
2025-12-04T09:37:02.3348594Z echo "Is keep-going label set? True"
2025-12-04T09:37:02.3348716Z 
2025-12-04T09:37:02.3348833Z echo
2025-12-04T09:37:02.3348929Z echo "Reenabled issues? "
2025-12-04T09:37:02.3352967Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T09:37:02.3353113Z env:
2025-12-04T09:37:02.3353203Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:37:02.3353335Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:37:02.3353508Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:37:02.3353671Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:37:02.3354053Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:37:02.3354418Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:37:02.3354535Z   AWS_REGION: us-east-1
2025-12-04T09:37:02.3354696Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:37:02.3354850Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:37:02.3356755Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:37:02.3356860Z ##[endgroup]
2025-12-04T09:37:02.3374304Z Filtered matrix:
2025-12-04T09:37:02.3378963Z {include: [{config: default, shard: 1, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, mem_leak_check: mem_leak_check}, {config: default, shard: 1, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, mem_leak_check: mem_leak_check, rerun_disabled_tests: rerun_disabled_tests}, {config: default, shard: 1, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests: rerun_disabled_tests, mem_leak_check: mem_leak_check}, {config: default, shard: 1, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests: rerun_disabled_tests}, {config: default, shard: 2, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, mem_leak_check: mem_leak_check}, {config: default, shard: 2, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, mem_leak_check: mem_leak_check, rerun_disabled_tests: rerun_disabled_tests}, {config: default, shard: 2, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests: rerun_disabled_tests, mem_leak_check: mem_leak_check}, {config: default, shard: 2, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests: rerun_disabled_tests}, {config: default, shard: 3, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, mem_leak_check: mem_leak_check}, {config: default, shard: 3, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, mem_leak_check: mem_leak_check, rerun_disabled_tests: rerun_disabled_tests}, {config: default, shard: 3, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests: rerun_disabled_tests, mem_leak_check: mem_leak_check}, {config: default, shard: 3, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests: rerun_disabled_tests}, {config: default, shard: 4, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, mem_leak_check: mem_leak_check}, {config: default, shard: 4, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, mem_leak_check: mem_leak_check, rerun_disabled_tests: rerun_disabled_tests}, {config: default, shard: 4, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests: rerun_disabled_tests, mem_leak_check: mem_leak_check}, {config: default, shard: 4, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests: rerun_disabled_tests}, {config: default, shard: 5, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, mem_leak_check: mem_leak_check}, {config: default, shard: 5, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, mem_leak_check: mem_leak_check, rerun_disabled_tests: rerun_disabled_tests}, {config: default, shard: 5, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests: rerun_disabled_tests, mem_leak_check: mem_leak_check}, {config: default, shard: 5, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests: rerun_disabled_tests}, {config: default, shard: 6, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, mem_leak_check: mem_leak_check}, {config: default, shard: 6, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, mem_leak_check: mem_leak_check, rerun_disabled_tests: rerun_disabled_tests}, {config: default, shard: 6, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests: rerun_disabled_tests, mem_leak_check: mem_leak_check}, {config: default, shard: 6, num_shards: 6, runner: linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests: rerun_disabled_tests}]}
2025-12-04T09:37:02.3383553Z 
2025-12-04T09:37:02.3383611Z Is the current job unstable? False
2025-12-04T09:37:02.3383708Z 
2025-12-04T09:37:02.3383764Z Is keep-going label set? True
2025-12-04T09:37:02.3383854Z 
2025-12-04T09:37:02.3383904Z Reenabled issues? 
2025-12-04T09:37:02.3409551Z ##[group]Run echo "timeout=$((JOB_TIMEOUT-30))" >> "${GITHUB_OUTPUT}"
2025-12-04T09:37:02.3409769Z echo "timeout=$((JOB_TIMEOUT-30))" >> "${GITHUB_OUTPUT}"
2025-12-04T09:37:02.3413940Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T09:37:02.3414093Z env:
2025-12-04T09:37:02.3414193Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:37:02.3414340Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:37:02.3414520Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:37:02.3414688Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:37:02.3415070Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:37:02.3415438Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:37:02.3415559Z   AWS_REGION: us-east-1
2025-12-04T09:37:02.3415777Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:37:02.3425686Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:37:02.3427579Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:37:02.3427689Z   JOB_TIMEOUT: 300
2025-12-04T09:37:02.3427785Z ##[endgroup]
2025-12-04T09:37:02.3471171Z ##[group]Run env | grep '^GITHUB' >> "/tmp/github_env_${GITHUB_RUN_ID}"
2025-12-04T09:37:02.3471456Z env | grep '^GITHUB' >> "/tmp/github_env_${GITHUB_RUN_ID}"
2025-12-04T09:37:02.3471706Z env | grep '^CI' >> "/tmp/github_env_${GITHUB_RUN_ID}"
2025-12-04T09:37:02.3476578Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T09:37:02.3476742Z env:
2025-12-04T09:37:02.3476847Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:37:02.3476993Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:37:02.3477182Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:37:02.3477360Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:37:02.3477761Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:37:02.3478189Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:37:02.3478320Z   AWS_REGION: us-east-1
2025-12-04T09:37:02.3478514Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:37:02.3478682Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:37:02.3480837Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:37:02.3480956Z ##[endgroup]
2025-12-04T09:37:02.3557112Z ##[group]Run set -x
2025-12-04T09:37:02.3557278Z set -x
2025-12-04T09:37:02.3557378Z 
2025-12-04T09:37:02.3557492Z if [[ $TEST_CONFIG == 'multigpu' ]]; then
2025-12-04T09:37:02.3557662Z   TEST_COMMAND=.ci/pytorch/multigpu-test.sh
2025-12-04T09:37:02.3557828Z elif [[ $BUILD_ENVIRONMENT == *onnx* ]]; then
2025-12-04T09:37:02.3557980Z   TEST_COMMAND=.ci/caffe2/test.sh
2025-12-04T09:37:02.3558108Z else
2025-12-04T09:37:02.3558220Z   TEST_COMMAND=.ci/pytorch/test.sh
2025-12-04T09:37:02.3558346Z fi
2025-12-04T09:37:02.3558439Z 
2025-12-04T09:37:02.3558577Z # detached container should get cleaned up by teardown_ec2_linux
2025-12-04T09:37:02.3558787Z # TODO: Stop building test binaries as part of the build phase
2025-12-04T09:37:02.3559121Z # Used for GPU_FLAG since that doesn't play nice
2025-12-04T09:37:02.3559311Z # shellcheck disable=SC2086,SC2090
2025-12-04T09:37:02.3559449Z container_name=$(docker run \
2025-12-04T09:37:02.3559580Z   ${GPU_FLAG:-} \
2025-12-04T09:37:02.3559703Z   -e BUILD_ENVIRONMENT \
2025-12-04T09:37:02.3559832Z   -e PR_NUMBER \
2025-12-04T09:37:02.3559951Z   -e GITHUB_ACTIONS \
2025-12-04T09:37:02.3560080Z   -e GITHUB_REPOSITORY \
2025-12-04T09:37:02.3560205Z   -e GITHUB_WORKFLOW \
2025-12-04T09:37:02.3560320Z   -e GITHUB_JOB \
2025-12-04T09:37:02.3560484Z   -e GITHUB_RUN_ID \
2025-12-04T09:37:02.3560602Z   -e GITHUB_RUN_NUMBER \
2025-12-04T09:37:02.3560724Z   -e GITHUB_RUN_ATTEMPT \
2025-12-04T09:37:02.3560846Z   -e JOB_ID \
2025-12-04T09:37:02.3560954Z   -e JOB_NAME \
2025-12-04T09:37:02.3561069Z   -e BASE_SHA \
2025-12-04T09:37:02.3561187Z   -e BRANCH \
2025-12-04T09:37:02.3561293Z   -e SHA1 \
2025-12-04T09:37:02.3561401Z   -e AWS_DEFAULT_REGION \
2025-12-04T09:37:02.3561559Z   -e IN_WHEEL_TEST \
2025-12-04T09:37:02.3561679Z   -e SHARD_NUMBER \
2025-12-04T09:37:02.3561796Z   -e TEST_CONFIG \
2025-12-04T09:37:02.3561916Z   -e NUM_TEST_SHARDS \
2025-12-04T09:37:02.3562042Z   -e REENABLED_ISSUES \
2025-12-04T09:37:02.3562171Z   -e CONTINUE_THROUGH_ERROR \
2025-12-04T09:37:02.3562301Z   -e VERBOSE_TEST_LOGS \
2025-12-04T09:37:02.3562425Z   -e TEST_SHOWLOCALS \
2025-12-04T09:37:02.3562550Z   -e NO_TEST_TIMEOUT \
2025-12-04T09:37:02.3562668Z   -e NO_TD \
2025-12-04T09:37:02.3562791Z   -e MAX_JOBS="$(nproc --ignore=2)" \
2025-12-04T09:37:02.3562943Z   -e PYTORCH_TEST_CUDA_MEM_LEAK_CHECK \
2025-12-04T09:37:02.3563091Z   -e PYTORCH_TEST_RERUN_DISABLED_TESTS \
2025-12-04T09:37:02.3563239Z   -e TESTS_TO_INCLUDE \
2025-12-04T09:37:02.3563368Z   -e HUGGING_FACE_HUB_TOKEN \
2025-12-04T09:37:02.3563501Z   -e DASHBOARD_TAG \
2025-12-04T09:37:02.3563660Z   --env-file="${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}" \
2025-12-04T09:37:02.3563828Z   --ulimit stack=10485760:83886080 \
2025-12-04T09:37:02.3563960Z   --ulimit core=0 \
2025-12-04T09:37:02.3564104Z   --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
2025-12-04T09:37:02.3564269Z   --security-opt seccomp=unconfined \
2025-12-04T09:37:02.3564411Z   --cap-add=SYS_PTRACE \
2025-12-04T09:37:02.3564540Z   --shm-size="8g" \
2025-12-04T09:37:02.3564655Z   --tty \
2025-12-04T09:37:02.3564765Z   --detach \
2025-12-04T09:37:02.3564884Z   --name="${container_name}" \
2025-12-04T09:37:02.3565018Z   --user jenkins \
2025-12-04T09:37:02.3565167Z   -v "${GITHUB_WORKSPACE}:/var/lib/jenkins/workspace" \
2025-12-04T09:37:02.3565338Z   -w /var/lib/jenkins/workspace \
2025-12-04T09:37:02.3565545Z   "${DOCKER_IMAGE}"
2025-12-04T09:37:02.3565660Z )
2025-12-04T09:37:02.3565773Z # save container name for later step
2025-12-04T09:37:02.3565944Z echo "CONTAINER_NAME=${container_name}" >> "$GITHUB_ENV"
2025-12-04T09:37:02.3566221Z # jenkins user does not have write permission to mounted workspace; work-around by copying within container to jenkins home
2025-12-04T09:37:02.3566573Z docker exec -t "${container_name}" sh -c "cd .. && cp -R workspace pytorch && cd pytorch && pip install dist/*.whl && ${TEST_COMMAND}"
2025-12-04T09:37:02.3570860Z shell: /usr/bin/bash -e {0}
2025-12-04T09:37:02.3570980Z env:
2025-12-04T09:37:02.3571082Z   GIT_DEFAULT_BRANCH: main
2025-12-04T09:37:02.3571225Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T09:37:02.3571411Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T09:37:02.3571636Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T09:37:02.3572027Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T09:37:02.3572403Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T09:37:02.3572527Z   AWS_REGION: us-east-1
2025-12-04T09:37:02.3572713Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T09:37:02.3572873Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T09:37:02.3574787Z   AWS_SESSION_TOKEN: ***
2025-12-04T09:37:02.3574925Z   BUILD_ENVIRONMENT: linux-noble-rocm-py3.12-mi300
2025-12-04T09:37:02.3575074Z   PR_NUMBER: 
2025-12-04T09:37:02.3575185Z   GITHUB_REPOSITORY: pytorch/pytorch
2025-12-04T09:37:02.3575319Z   GITHUB_WORKFLOW: rocm-mi300
2025-12-04T09:37:02.3575439Z   GITHUB_JOB: test
2025-12-04T09:37:02.3575550Z   GITHUB_RUN_ID: 19922812470
2025-12-04T09:37:02.3575675Z   GITHUB_RUN_NUMBER: 14122
2025-12-04T09:37:02.3575793Z   GITHUB_RUN_ATTEMPT: 1
2025-12-04T09:37:02.3575903Z   JOB_ID: 57116139325
2025-12-04T09:37:02.3576114Z   JOB_NAME: linux-noble-rocm-py3.12-mi300 / test (default, 3, 6, linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests)
2025-12-04T09:37:02.3576333Z   BRANCH: main
2025-12-04T09:37:02.3576452Z   SHA1: ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T09:37:02.3576613Z   BASE_SHA: ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T09:37:02.3576758Z   TEST_CONFIG: default
2025-12-04T09:37:02.3576868Z   SHARD_NUMBER: 3
2025-12-04T09:37:02.3576970Z   NUM_TEST_SHARDS: 6
2025-12-04T09:37:02.3577079Z   REENABLED_ISSUES: 
2025-12-04T09:37:02.3577194Z   CONTINUE_THROUGH_ERROR: True
2025-12-04T09:37:02.3577318Z   VERBOSE_TEST_LOGS: False
2025-12-04T09:37:02.3577435Z   TEST_SHOWLOCALS: False
2025-12-04T09:37:02.3577549Z   NO_TEST_TIMEOUT: False
2025-12-04T09:37:02.3577661Z   NO_TD: False
2025-12-04T09:37:02.3577934Z   DOCKER_IMAGE: 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T09:37:02.3578245Z   PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: 0
2025-12-04T09:37:02.3578384Z   PYTORCH_TEST_RERUN_DISABLED_TESTS: 1
2025-12-04T09:37:02.3578516Z   TESTS_TO_INCLUDE: 
2025-12-04T09:37:02.3578623Z   DASHBOARD_TAG: 
2025-12-04T09:37:02.3578769Z   HUGGING_FACE_HUB_TOKEN: ***
2025-12-04T09:37:02.3578890Z ##[endgroup]
2025-12-04T09:37:02.3595236Z + [[ default == \m\u\l\t\i\g\p\u ]]
2025-12-04T09:37:02.3595391Z + [[ linux-noble-rocm-py3.12-mi300 == *onnx* ]]
2025-12-04T09:37:02.3595549Z + TEST_COMMAND=.ci/pytorch/test.sh
2025-12-04T09:37:02.3601208Z +++ nproc --ignore=2
2025-12-04T09:37:02.3610154Z ++ docker run --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host -e BUILD_ENVIRONMENT -e PR_NUMBER -e GITHUB_ACTIONS -e GITHUB_REPOSITORY -e GITHUB_WORKFLOW -e GITHUB_JOB -e GITHUB_RUN_ID -e GITHUB_RUN_NUMBER -e GITHUB_RUN_ATTEMPT -e JOB_ID -e JOB_NAME -e BASE_SHA -e BRANCH -e SHA1 -e AWS_DEFAULT_REGION -e IN_WHEEL_TEST -e SHARD_NUMBER -e TEST_CONFIG -e NUM_TEST_SHARDS -e REENABLED_ISSUES -e CONTINUE_THROUGH_ERROR -e VERBOSE_TEST_LOGS -e TEST_SHOWLOCALS -e NO_TEST_TIMEOUT -e NO_TD -e MAX_JOBS=126 -e PYTORCH_TEST_CUDA_MEM_LEAK_CHECK -e PYTORCH_TEST_RERUN_DISABLED_TESTS -e TESTS_TO_INCLUDE -e HUGGING_FACE_HUB_TOKEN -e DASHBOARD_TAG --env-file=/home/runner/_work/_temp/github_env_19922812470 --ulimit stack=10485760:83886080 --ulimit core=0 --env-file=/tmp/github_env_19922812470 --security-opt seccomp=unconfined --cap-add=SYS_PTRACE --shm-size=8g --tty --detach --name= --user jenkins -v /home/runner/_work/pytorch/pytorch:/var/lib/jenkins/workspace -w /var/lib/jenkins/workspace 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/ci-image:pytorch-linux-noble-rocm-n-py3-f0cd68561080d537ef3d3d6f81b25a6416ad600a
2025-12-04T09:37:04.7069947Z + container_name=258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19
2025-12-04T09:37:04.7070357Z + echo CONTAINER_NAME=258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19
2025-12-04T09:37:04.7071342Z + docker exec -t 258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19 sh -c 'cd .. && cp -R workspace pytorch && cd pytorch && pip install dist/*.whl && .ci/pytorch/test.sh'
2025-12-04T09:37:08.0429958Z Processing ./dist/torch-2.10.0a0+gitffd9b0f-cp312-cp312-linux_x86_64.whl
2025-12-04T09:37:08.5780298Z Requirement already satisfied: filelock in /opt/conda/envs/py_3.12/lib/python3.12/site-packages (from torch==2.10.0a0+gitffd9b0f) (3.18.0)
2025-12-04T09:37:08.5780852Z Requirement already satisfied: typing-extensions>=4.10.0 in /opt/conda/envs/py_3.12/lib/python3.12/site-packages (from torch==2.10.0a0+gitffd9b0f) (4.12.2)
2025-12-04T09:37:08.5782592Z Requirement already satisfied: setuptools in /opt/conda/envs/py_3.12/lib/python3.12/site-packages (from torch==2.10.0a0+gitffd9b0f) (78.1.1)
2025-12-04T09:37:08.5783059Z Requirement already satisfied: sympy>=1.13.3 in /opt/conda/envs/py_3.12/lib/python3.12/site-packages (from torch==2.10.0a0+gitffd9b0f) (1.13.3)
2025-12-04T09:37:08.5784991Z Requirement already satisfied: networkx>=2.5.1 in /opt/conda/envs/py_3.12/lib/python3.12/site-packages (from torch==2.10.0a0+gitffd9b0f) (2.8.8)
2025-12-04T09:37:08.5785387Z Requirement already satisfied: jinja2 in /opt/conda/envs/py_3.12/lib/python3.12/site-packages (from torch==2.10.0a0+gitffd9b0f) (3.1.6)
2025-12-04T09:37:08.5787131Z Requirement already satisfied: fsspec>=0.8.5 in /opt/conda/envs/py_3.12/lib/python3.12/site-packages (from torch==2.10.0a0+gitffd9b0f) (2025.10.0)
2025-12-04T09:37:08.5833551Z Requirement already satisfied: mpmath<1.4,>=1.1.0 in /opt/conda/envs/py_3.12/lib/python3.12/site-packages (from sympy>=1.13.3->torch==2.10.0a0+gitffd9b0f) (1.3.0)
2025-12-04T09:37:08.5852805Z Requirement already satisfied: MarkupSafe>=2.0 in /opt/conda/envs/py_3.12/lib/python3.12/site-packages (from jinja2->torch==2.10.0a0+gitffd9b0f) (3.0.3)
2025-12-04T09:37:08.7199912Z Installing collected packages: torch
2025-12-04T09:37:14.5480145Z Successfully installed torch-2.10.0a0+gitffd9b0f
2025-12-04T09:37:14.5941998Z + export TERM=vt100
2025-12-04T09:37:14.5942361Z + TERM=vt100
2025-12-04T09:37:14.5945868Z ++ dirname .ci/pytorch/test.sh
2025-12-04T09:37:14.5955681Z + source .ci/pytorch/common.sh
2025-12-04T09:37:14.5961572Z +++ dirname .ci/pytorch/common.sh
2025-12-04T09:37:14.5972591Z ++ source .ci/pytorch/common_utils.sh
2025-12-04T09:37:14.5974319Z +++ declare -f -t trap_add
2025-12-04T09:37:14.5979400Z ++ set -ex -o pipefail
2025-12-04T09:37:14.5979539Z ++ [[ linux-noble-rocm-py3.12-mi300 == *rocm* ]]
2025-12-04T09:37:14.5979682Z ++ unset HIP_PLATFORM
2025-12-04T09:37:14.5979799Z ++ export PYTORCH_TEST_WITH_ROCM=1
2025-12-04T09:37:14.5979921Z ++ PYTORCH_TEST_WITH_ROCM=1
2025-12-04T09:37:14.5980036Z ++ BUILD_TEST_LIBTORCH=0
2025-12-04T09:37:14.5985862Z ++ dirname .ci/pytorch/test.sh
2025-12-04T09:37:14.5996368Z + source .ci/pytorch/common-build.sh
2025-12-04T09:37:14.5998132Z ++ [[ linux-noble-rocm-py3.12-mi300 != *win-* ]]
2025-12-04T09:37:14.6004205Z ++++ dirname .ci/pytorch/common-build.sh
2025-12-04T09:37:14.6009154Z +++ cd .ci/pytorch
2025-12-04T09:37:14.6009496Z +++ pwd -P
2025-12-04T09:37:14.6011982Z ++ script_dir=/var/lib/jenkins/pytorch/.ci/pytorch
2025-12-04T09:37:14.6012188Z ++ [[ linux-noble-rocm-py3.12-mi300 == *-pch* ]]
2025-12-04T09:37:14.6012314Z ++ which sccache
2025-12-04T09:37:14.6028380Z ++ [[ -z '' ]]
2025-12-04T09:37:14.6028495Z ++ unset SCCACHE_BUCKET
2025-12-04T09:37:14.6029480Z ++ unset SCCACHE_REGION
2025-12-04T09:37:14.6029852Z ++ sccache --stop-server
2025-12-04T09:37:14.6051805Z ++ true
2025-12-04T09:37:14.6052033Z ++ rm -f /var/lib/jenkins/sccache_error.log
2025-12-04T09:37:14.6063103Z ++ trap_add sccache_epilogue EXIT
2025-12-04T09:37:14.6063371Z ++ trap_add_cmd=sccache_epilogue
2025-12-04T09:37:14.6063592Z ++ shift
2025-12-04T09:37:14.6063782Z ++ for trap_add_name in "$@"
2025-12-04T09:37:14.6069292Z ++++ trap -p EXIT
2025-12-04T09:37:14.6071081Z +++ eval 'extract_trap_cmd '
2025-12-04T09:37:14.6071309Z ++++ extract_trap_cmd
2025-12-04T09:37:14.6071504Z ++++ printf '%s\n' ''
2025-12-04T09:37:14.6071707Z +++ printf '%s\n' sccache_epilogue
2025-12-04T09:37:14.6073162Z ++ trap -- '
2025-12-04T09:37:14.6073339Z sccache_epilogue' EXIT
2025-12-04T09:37:14.6073536Z ++ [[ -n '' ]]
2025-12-04T09:37:14.6073746Z ++ [[ linux-noble-rocm-py3.12-mi300 == *rocm* ]]
2025-12-04T09:37:14.6074044Z ++ SCCACHE_ERROR_LOG=/var/lib/jenkins/sccache_error.log
2025-12-04T09:37:14.6074310Z ++ SCCACHE_IDLE_TIMEOUT=0
2025-12-04T09:37:14.6074518Z ++ sccache --start-server
2025-12-04T09:37:14.6091747Z sccache: Starting the server...
2025-12-04T09:37:14.6293300Z sccache: Listening on address 127.0.0.1:4226
2025-12-04T09:37:14.6304178Z ++ sccache --zero-stats
2025-12-04T09:37:14.6316638Z Statistics zeroed.
2025-12-04T09:37:14.6318847Z ++ which ccache
2025-12-04T09:37:14.6327737Z + [[ linux-noble-rocm-py3.12-mi300 != *rocm* ]]
2025-12-04T09:37:14.6328079Z + [[ linux-noble-rocm-py3.12-mi300 == *cuda* ]]
2025-12-04T09:37:14.6328359Z + echo 'Environment variables:'
2025-12-04T09:37:14.6328587Z Environment variables:
2025-12-04T09:37:14.6328781Z + env
2025-12-04T09:37:14.6334244Z GITHUB_WORKSPACE=/home/runner/_work/pytorch/pytorch
2025-12-04T09:37:14.6334575Z CONTINUE_THROUGH_ERROR=True
2025-12-04T09:37:14.6334840Z BUILD_ENVIRONMENT=linux-noble-rocm-py3.12-mi300
2025-12-04T09:37:14.6335166Z HOSTNAME=linux.rocm.gpu.gfx942.1.b-gwk9b-runner-xf6tf
2025-12-04T09:37:14.6335613Z GITHUB_PATH=/home/runner/_work/_temp/_runner_file_commands/add_path_8f05a0c2-1271-4052-93aa-2bb5fce40c1a
2025-12-04T09:37:14.6335979Z GITHUB_ACTION=__run_2
2025-12-04T09:37:14.6336179Z PYTORCH_TEST_CUDA_MEM_LEAK_CHECK=0
2025-12-04T09:37:14.6336398Z GITHUB_RUN_NUMBER=14122
2025-12-04T09:37:14.6336580Z TEST_CONFIG=default
2025-12-04T09:37:14.6336835Z RUNNER_NAME=linux.rocm.gpu.gfx942.1.b-gwk9b-runner-xf6tf
2025-12-04T09:37:14.6337111Z GITHUB_REPOSITORY_OWNER_ID=21003710
2025-12-04T09:37:14.6337341Z AWS_DEFAULT_REGION=us-east-1
2025-12-04T09:37:14.6337596Z RUNNER_ARTIFACT_DIR=/home/runner/_work/_temp/artifacts
2025-12-04T09:37:14.6337860Z GITHUB_TRIGGERING_ACTOR=pytorchmergebot
2025-12-04T09:37:14.6338085Z GITHUB_REF_TYPE=branch
2025-12-04T09:37:14.6338303Z BASE_SHA=ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T09:37:14.6338698Z HUGGING_FACE_HUB_TOKEN=***
2025-12-04T09:37:14.6340826Z ***
2025-12-04T09:37:14.6340996Z GITHUB_REPOSITORY_ID=65600975
2025-12-04T09:37:14.6341195Z GITHUB_ACTIONS=true
2025-12-04T09:37:14.6341398Z SHA1=ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T09:37:14.6341661Z GITHUB_SHA=ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T09:37:14.6342023Z GITHUB_WORKFLOW_REF=pytorch/pytorch/.github/workflows/rocm-mi300.yml@refs/heads/main
2025-12-04T09:37:14.6342346Z UCC_HOME=/usr
2025-12-04T09:37:14.6342519Z RUNNER_ENVIRONMENT=self-hosted
2025-12-04T09:37:14.6342717Z VERBOSE_TEST_LOGS=False
2025-12-04T09:37:14.6342908Z GITHUB_REF=refs/heads/main
2025-12-04T09:37:14.6343100Z RUNNER_OS=Linux
2025-12-04T09:37:14.6343262Z SHARD_NUMBER=3
2025-12-04T09:37:14.6343693Z GITHUB_REF_PROTECTED=true
2025-12-04T09:37:14.6343891Z RUNNER_MANUALLY_TRAP_SIG=1
2025-12-04T09:37:14.6344075Z HOME=/var/lib/jenkins
2025-12-04T09:37:14.6344280Z GITHUB_API_URL=https://api.github.com
2025-12-04T09:37:14.6344511Z PYTORCH_TEST_RERUN_DISABLED_TESTS=1
2025-12-04T09:37:14.6344758Z RUNNER_DOCS_DIR=/home/runner/_work/_temp/docs
2025-12-04T09:37:14.6344975Z LANG=C.UTF-8
2025-12-04T09:37:14.6345173Z UCX_COMMIT=29831d319e6be55cb8c768ca61de335c934ca39e
2025-12-04T09:37:14.6345424Z PYTORCH_TEST_WITH_ROCM=1
2025-12-04T09:37:14.6345627Z RUNNER_TRACKING_ID=github_92f77cd4-044a-4aa5-8af7-1de94326986d
2025-12-04T09:37:14.6345837Z RUNNER_ARCH=X64
2025-12-04T09:37:14.6345987Z RUNNER_TEMP=/home/runner/_work/_temp
2025-12-04T09:37:14.6346151Z NUM_TEST_SHARDS=6
2025-12-04T09:37:14.6346288Z UCX_HOME=/usr
2025-12-04T09:37:14.6346546Z GITHUB_STATE=/home/runner/_work/_temp/_runner_file_commands/save_state_8f05a0c2-1271-4052-93aa-2bb5fce40c1a
2025-12-04T09:37:14.6347028Z JOB_NAME=linux-noble-rocm-py3.12-mi300 / test (default, 3, 6, linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests)
2025-12-04T09:37:14.6347335Z MAGMA_HOME=/opt/rocm/magma
2025-12-04T09:37:14.6347603Z GITHUB_ENV=/home/runner/_work/_temp/_runner_file_commands/set_env_8f05a0c2-1271-4052-93aa-2bb5fce40c1a
2025-12-04T09:37:14.6348058Z GITHUB_EVENT_PATH=/home/runner/_work/_temp/_github_workflow/event.json
2025-12-04T09:37:14.6348297Z GITHUB_EVENT_NAME=schedule
2025-12-04T09:37:14.6348518Z GITHUB_ACTIONS_RUNNER_EXTRA_USER_AGENT=actions-runner-controller/0.12.1
2025-12-04T09:37:14.6348756Z DASHBOARD_TAG=
2025-12-04T09:37:14.6348890Z GITHUB_RUN_ID=19922812470
2025-12-04T09:37:14.6349183Z GITHUB_STEP_SUMMARY=/home/runner/_work/_temp/_runner_file_commands/step_summary_8f05a0c2-1271-4052-93aa-2bb5fce40c1a
2025-12-04T09:37:14.6349496Z GITHUB_ACTOR=pytorchmergebot
2025-12-04T09:37:14.6349656Z PR_NUMBER=
2025-12-04T09:37:14.6349783Z GITHUB_RUN_ATTEMPT=1
2025-12-04T09:37:14.6349938Z ANACONDA_PYTHON_VERSION=3.12
2025-12-04T09:37:14.6350126Z GITHUB_GRAPHQL_URL=https://api.github.com/graphql
2025-12-04T09:37:14.6350317Z TERM=vt100
2025-12-04T09:37:14.6350497Z INSTALLED_VISION=yes
2025-12-04T09:37:14.6350637Z BRANCH=main
2025-12-04T09:37:14.6350771Z OPENSSL_ROOT_DIR=/opt/openssl
2025-12-04T09:37:14.6350931Z TESTS_TO_INCLUDE=
2025-12-04T09:37:14.6351175Z GITHUB_ACTION_PATH=/home/runner/_work/pytorch/pytorch/./.github/actions/setup-rocm
2025-12-04T09:37:14.6351439Z GITHUB_SERVER_URL=https://github.com
2025-12-04T09:37:14.6351633Z PYTORCH_ROCM_ARCH=gfx90a;gfx942;gfx950;gfx1100
2025-12-04T09:37:14.6351841Z UCC_COMMIT=9f4b242cbbd8b1462cbc732eb29316cdfa124b77
2025-12-04T09:37:14.6352033Z REENABLED_ISSUES=
2025-12-04T09:37:14.6352160Z SHLVL=1
2025-12-04T09:37:14.6352288Z MAX_JOBS=126
2025-12-04T09:37:14.6352465Z RUNNER_TEST_RESULTS_DIR=/home/runner/_work/_temp/test-results
2025-12-04T09:37:14.6352686Z GITHUB_ACTOR_ID=97764156
2025-12-04T09:37:14.6352851Z RUNNER_TOOL_CACHE=/home/runner/_work/_tool
2025-12-04T09:37:14.6353072Z GITHUB_WORKFLOW_SHA=ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T09:37:14.6353286Z GITHUB_REF_NAME=main
2025-12-04T09:37:14.6353432Z ROCM_PATH=/opt/rocm
2025-12-04T09:37:14.6353574Z GITHUB_JOB=test
2025-12-04T09:37:14.6353709Z NO_TEST_TIMEOUT=False
2025-12-04T09:37:14.6353864Z GITHUB_REPOSITORY=pytorch/pytorch
2025-12-04T09:37:14.6354031Z LC_ALL=C.UTF-8
2025-12-04T09:37:14.6354160Z GITHUB_RETENTION_DAYS=90
2025-12-04T09:37:14.6354323Z RUNNER_WORKSPACE=/home/runner/_work/pytorch
2025-12-04T09:37:14.6354505Z OPENSSL_DIR=/opt/openssl
2025-12-04T09:37:14.6354660Z GITHUB_ACTION_REPOSITORY=
2025-12-04T09:37:14.6355157Z PATH=/opt/cache/bin:/opt/rocm/llvm/bin:/opt/rocm/opencl/bin:/opt/rocm/hip/bin:/opt/rocm/hcc/bin:/opt/rocm/bin:/opt/conda/envs/py_3.12/bin:/opt/conda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
2025-12-04T09:37:14.6355641Z GITHUB_BASE_REF=
2025-12-04T09:37:14.6355748Z CI=true
2025-12-04T09:37:14.6355856Z GITHUB_REPOSITORY_OWNER=pytorch
2025-12-04T09:37:14.6355987Z JOB_ID=57116139325
2025-12-04T09:37:14.6356095Z GITHUB_HEAD_REF=
2025-12-04T09:37:14.6356288Z GITHUB_ACTION_REF=
2025-12-04T09:37:14.6356394Z TEST_SHOWLOCALS=False
2025-12-04T09:37:14.6356515Z GITHUB_WORKFLOW=rocm-mi300
2025-12-04T09:37:14.6356644Z DEBIAN_FRONTEND=noninteractive
2025-12-04T09:37:14.6356872Z GITHUB_OUTPUT=/home/runner/_work/_temp/_runner_file_commands/set_output_8f05a0c2-1271-4052-93aa-2bb5fce40c1a
2025-12-04T09:37:14.6357104Z NO_TD=False
2025-12-04T09:37:14.6357209Z OLDPWD=/var/lib/jenkins
2025-12-04T09:37:14.6357324Z _=/usr/bin/env
2025-12-04T09:37:14.6357471Z ++ python -c 'import site; print(site.getsitepackages()[0])'
2025-12-04T09:37:14.6397878Z + TORCH_INSTALL_DIR=/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch
2025-12-04T09:37:14.6398228Z + TORCH_BIN_DIR=/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/bin
2025-12-04T09:37:14.6398537Z + TORCH_LIB_DIR=/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/lib
2025-12-04T09:37:14.6399665Z + TORCH_TEST_DIR=/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/test
2025-12-04T09:37:14.6399906Z + BUILD_DIR=build
2025-12-04T09:37:14.6400059Z + BUILD_RENAMED_DIR=build_renamed
2025-12-04T09:37:14.6400225Z + BUILD_BIN_DIR=build/bin
2025-12-04T09:37:14.6400373Z + SHARD_NUMBER=3
2025-12-04T09:37:14.6400560Z + NUM_TEST_SHARDS=6
2025-12-04T09:37:14.6400720Z + export TORCH_SERIALIZATION_DEBUG=1
2025-12-04T09:37:14.6400896Z + TORCH_SERIALIZATION_DEBUG=1
2025-12-04T09:37:14.6401055Z + export VALGRIND=ON
2025-12-04T09:37:14.6401197Z + VALGRIND=ON
2025-12-04T09:37:14.6401352Z + [[ linux-noble-rocm-py3.12-mi300 == *clang9* ]]
2025-12-04T09:37:14.6401555Z + [[ linux-noble-rocm-py3.12-mi300 == *xpu* ]]
2025-12-04T09:37:14.6401729Z + detect_cuda_arch
2025-12-04T09:37:14.6401907Z + [[ linux-noble-rocm-py3.12-mi300 == *cuda* ]]
2025-12-04T09:37:14.6402109Z + [[ linux-noble-rocm-py3.12-mi300 == *s390x* ]]
2025-12-04T09:37:14.6402282Z + [[ 1 == \1 ]]
2025-12-04T09:37:14.6402425Z + ulimit -c 0
2025-12-04T09:37:14.6402569Z + [[ linux-noble-rocm-py3.12-mi300 != *bazel* ]]
2025-12-04T09:37:14.6402754Z ++ realpath build/custom_test_artifacts
2025-12-04T09:37:14.6408188Z + CUSTOM_TEST_ARTIFACT_BUILD_DIR=/var/lib/jenkins/pytorch/build/custom_test_artifacts
2025-12-04T09:37:14.6408726Z + [[ -n '' ]]
2025-12-04T09:37:14.6408846Z + echo 'Environment variables'
2025-12-04T09:37:14.6408995Z Environment variables
2025-12-04T09:37:14.6409110Z + env
2025-12-04T09:37:14.6417367Z GITHUB_WORKSPACE=/home/runner/_work/pytorch/pytorch
2025-12-04T09:37:14.6417531Z CONTINUE_THROUGH_ERROR=True
2025-12-04T09:37:14.6417683Z BUILD_ENVIRONMENT=linux-noble-rocm-py3.12-mi300
2025-12-04T09:37:14.6417897Z HOSTNAME=linux.rocm.gpu.gfx942.1.b-gwk9b-runner-xf6tf
2025-12-04T09:37:14.6418166Z GITHUB_PATH=/home/runner/_work/_temp/_runner_file_commands/add_path_8f05a0c2-1271-4052-93aa-2bb5fce40c1a
2025-12-04T09:37:14.6418400Z GITHUB_ACTION=__run_2
2025-12-04T09:37:14.6418519Z PYTORCH_TEST_CUDA_MEM_LEAK_CHECK=0
2025-12-04T09:37:14.6418664Z GITHUB_RUN_NUMBER=14122
2025-12-04T09:37:14.6418778Z TEST_CONFIG=default
2025-12-04T09:37:14.6418928Z RUNNER_NAME=linux.rocm.gpu.gfx942.1.b-gwk9b-runner-xf6tf
2025-12-04T09:37:14.6419099Z GITHUB_REPOSITORY_OWNER_ID=21003710
2025-12-04T09:37:14.6419242Z AWS_DEFAULT_REGION=us-east-1
2025-12-04T09:37:14.6419397Z RUNNER_ARTIFACT_DIR=/home/runner/_work/_temp/artifacts
2025-12-04T09:37:14.6419559Z GITHUB_TRIGGERING_ACTOR=pytorchmergebot
2025-12-04T09:37:14.6419700Z GITHUB_REF_TYPE=branch
2025-12-04T09:37:14.6419839Z BASE_SHA=ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T09:37:14.6420067Z HUGGING_FACE_HUB_TOKEN=***
2025-12-04T09:37:14.6420213Z ***
2025-12-04T09:37:14.6420315Z GITHUB_REPOSITORY_ID=65600975
2025-12-04T09:37:14.6420485Z GITHUB_ACTIONS=true
2025-12-04T09:37:14.6420616Z SHA1=ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T09:37:14.6420778Z GITHUB_SHA=ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T09:37:14.6421008Z GITHUB_WORKFLOW_REF=pytorch/pytorch/.github/workflows/rocm-mi300.yml@refs/heads/main
2025-12-04T09:37:14.6421214Z UCC_HOME=/usr
2025-12-04T09:37:14.6421324Z TORCH_SERIALIZATION_DEBUG=1
2025-12-04T09:37:14.6421513Z RUNNER_ENVIRONMENT=self-hosted
2025-12-04T09:37:14.6421642Z VERBOSE_TEST_LOGS=False
2025-12-04T09:37:14.6421758Z GITHUB_REF=refs/heads/main
2025-12-04T09:37:14.6421877Z RUNNER_OS=Linux
2025-12-04T09:37:14.6421981Z SHARD_NUMBER=3
2025-12-04T09:37:14.6422089Z GITHUB_REF_PROTECTED=true
2025-12-04T09:37:14.6422210Z RUNNER_MANUALLY_TRAP_SIG=1
2025-12-04T09:37:14.6422331Z HOME=/var/lib/jenkins
2025-12-04T09:37:14.6422465Z GITHUB_API_URL=https://api.github.com
2025-12-04T09:37:14.6422615Z PYTORCH_TEST_RERUN_DISABLED_TESTS=1
2025-12-04T09:37:14.6422762Z RUNNER_DOCS_DIR=/home/runner/_work/_temp/docs
2025-12-04T09:37:14.6422899Z LANG=C.UTF-8
2025-12-04T09:37:14.6423022Z UCX_COMMIT=29831d319e6be55cb8c768ca61de335c934ca39e
2025-12-04T09:37:14.6423176Z PYTORCH_TEST_WITH_ROCM=1
2025-12-04T09:37:14.6423337Z RUNNER_TRACKING_ID=github_92f77cd4-044a-4aa5-8af7-1de94326986d
2025-12-04T09:37:14.6423548Z RUNNER_ARCH=X64
2025-12-04T09:37:14.6423661Z RUNNER_TEMP=/home/runner/_work/_temp
2025-12-04T09:37:14.6423797Z NUM_TEST_SHARDS=6
2025-12-04T09:37:14.6423901Z UCX_HOME=/usr
2025-12-04T09:37:14.6424110Z GITHUB_STATE=/home/runner/_work/_temp/_runner_file_commands/save_state_8f05a0c2-1271-4052-93aa-2bb5fce40c1a
2025-12-04T09:37:14.6424456Z JOB_NAME=linux-noble-rocm-py3.12-mi300 / test (default, 3, 6, linux.rocm.gpu.gfx942.1.b, rerun_disabled_tests)
2025-12-04T09:37:14.6424697Z MAGMA_HOME=/opt/rocm/magma
2025-12-04T09:37:14.6424907Z GITHUB_ENV=/home/runner/_work/_temp/_runner_file_commands/set_env_8f05a0c2-1271-4052-93aa-2bb5fce40c1a
2025-12-04T09:37:14.6425179Z GITHUB_EVENT_PATH=/home/runner/_work/_temp/_github_workflow/event.json
2025-12-04T09:37:14.6425445Z GITHUB_EVENT_NAME=schedule
2025-12-04T09:37:14.6425609Z GITHUB_ACTIONS_RUNNER_EXTRA_USER_AGENT=actions-runner-controller/0.12.1
2025-12-04T09:37:14.6425802Z DASHBOARD_TAG=
2025-12-04T09:37:14.6425902Z GITHUB_RUN_ID=19922812470
2025-12-04T09:37:14.6426114Z GITHUB_STEP_SUMMARY=/home/runner/_work/_temp/_runner_file_commands/step_summary_8f05a0c2-1271-4052-93aa-2bb5fce40c1a
2025-12-04T09:37:14.6426346Z GITHUB_ACTOR=pytorchmergebot
2025-12-04T09:37:14.6426457Z PR_NUMBER=
2025-12-04T09:37:14.6426549Z GITHUB_RUN_ATTEMPT=1
2025-12-04T09:37:14.6426649Z VALGRIND=ON
2025-12-04T09:37:14.6426746Z ANACONDA_PYTHON_VERSION=3.12
2025-12-04T09:37:14.6426882Z GITHUB_GRAPHQL_URL=https://api.github.com/graphql
2025-12-04T09:37:14.6427020Z TERM=vt100
2025-12-04T09:37:14.6427110Z INSTALLED_VISION=yes
2025-12-04T09:37:14.6427209Z BRANCH=main
2025-12-04T09:37:14.6427303Z OPENSSL_ROOT_DIR=/opt/openssl
2025-12-04T09:37:14.6427415Z TESTS_TO_INCLUDE=
2025-12-04T09:37:14.6427576Z GITHUB_ACTION_PATH=/home/runner/_work/pytorch/pytorch/./.github/actions/setup-rocm
2025-12-04T09:37:14.6427767Z GITHUB_SERVER_URL=https://github.com
2025-12-04T09:37:14.6427906Z PYTORCH_ROCM_ARCH=gfx90a;gfx942;gfx950;gfx1100
2025-12-04T09:37:14.6428057Z UCC_COMMIT=9f4b242cbbd8b1462cbc732eb29316cdfa124b77
2025-12-04T09:37:14.6428197Z REENABLED_ISSUES=
2025-12-04T09:37:14.6428292Z SHLVL=1
2025-12-04T09:37:14.6428378Z MAX_JOBS=126
2025-12-04T09:37:14.6428505Z RUNNER_TEST_RESULTS_DIR=/home/runner/_work/_temp/test-results
2025-12-04T09:37:14.6428657Z GITHUB_ACTOR_ID=97764156
2025-12-04T09:37:14.6428773Z RUNNER_TOOL_CACHE=/home/runner/_work/_tool
2025-12-04T09:37:14.6428932Z GITHUB_WORKFLOW_SHA=ffd9b0fb4355e97af82fc42cf185c3ffa0fc0a32
2025-12-04T09:37:14.6429082Z GITHUB_REF_NAME=main
2025-12-04T09:37:14.6429183Z ROCM_PATH=/opt/rocm
2025-12-04T09:37:14.6429279Z GITHUB_JOB=test
2025-12-04T09:37:14.6429376Z NO_TEST_TIMEOUT=False
2025-12-04T09:37:14.6429488Z GITHUB_REPOSITORY=pytorch/pytorch
2025-12-04T09:37:14.6429606Z LC_ALL=C.UTF-8
2025-12-04T09:37:14.6429706Z GITHUB_RETENTION_DAYS=90
2025-12-04T09:37:14.6429824Z RUNNER_WORKSPACE=/home/runner/_work/pytorch
2025-12-04T09:37:14.6429957Z OPENSSL_DIR=/opt/openssl
2025-12-04T09:37:14.6430069Z GITHUB_ACTION_REPOSITORY=
2025-12-04T09:37:14.6430535Z PATH=/opt/cache/bin:/opt/rocm/llvm/bin:/opt/rocm/opencl/bin:/opt/rocm/hip/bin:/opt/rocm/hcc/bin:/opt/rocm/bin:/opt/conda/envs/py_3.12/bin:/opt/conda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
2025-12-04T09:37:14.6430898Z GITHUB_BASE_REF=
2025-12-04T09:37:14.6430994Z CI=true
2025-12-04T09:37:14.6431089Z GITHUB_REPOSITORY_OWNER=pytorch
2025-12-04T09:37:14.6431204Z JOB_ID=57116139325
2025-12-04T09:37:14.6431299Z GITHUB_HEAD_REF=
2025-12-04T09:37:14.6431396Z GITHUB_ACTION_REF=
2025-12-04T09:37:14.6431495Z TEST_SHOWLOCALS=False
2025-12-04T09:37:14.6431601Z GITHUB_WORKFLOW=rocm-mi300
2025-12-04T09:37:14.6431720Z DEBIAN_FRONTEND=noninteractive
2025-12-04T09:37:14.6431925Z GITHUB_OUTPUT=/home/runner/_work/_temp/_runner_file_commands/set_output_8f05a0c2-1271-4052-93aa-2bb5fce40c1a
2025-12-04T09:37:14.6432135Z NO_TD=False
2025-12-04T09:37:14.6432226Z OLDPWD=/var/lib/jenkins
2025-12-04T09:37:14.6432328Z _=/usr/bin/env
2025-12-04T09:37:14.6432424Z + echo 'Testing pytorch'
2025-12-04T09:37:14.6432530Z Testing pytorch
2025-12-04T09:37:14.6432694Z + export LANG=C.UTF-8
2025-12-04T09:37:14.6432791Z + LANG=C.UTF-8
2025-12-04T09:37:14.6432882Z + PR_NUMBER=
2025-12-04T09:37:14.6432987Z + [[ default == \d\e\f\a\u\l\t ]]
2025-12-04T09:37:14.6433109Z + export CUDA_VISIBLE_DEVICES=0
2025-12-04T09:37:14.6433225Z + CUDA_VISIBLE_DEVICES=0
2025-12-04T09:37:14.6433336Z + export HIP_VISIBLE_DEVICES=0
2025-12-04T09:37:14.6433456Z + HIP_VISIBLE_DEVICES=0
2025-12-04T09:37:14.6433568Z + [[ default == \d\i\s\t\r\i\b\u\t\e\d ]]
2025-12-04T09:37:14.6433691Z + [[ default == \s\l\o\w ]]
2025-12-04T09:37:14.6433830Z + [[ linux-noble-rocm-py3.12-mi300 == *slow-gradcheck* ]]
2025-12-04T09:37:14.6433991Z + [[ linux-noble-rocm-py3.12-mi300 == *cuda* ]]
2025-12-04T09:37:14.6434136Z + [[ linux-noble-rocm-py3.12-mi300 == *rocm* ]]
2025-12-04T09:37:14.6434281Z + export PYTORCH_TESTING_DEVICE_ONLY_FOR=cuda
2025-12-04T09:37:14.6434420Z + PYTORCH_TESTING_DEVICE_ONLY_FOR=cuda
2025-12-04T09:37:14.6434547Z + [[ default == *crossref* ]]
2025-12-04T09:37:14.6434677Z + [[ linux-noble-rocm-py3.12-mi300 == *rocm* ]]
2025-12-04T09:37:14.6434808Z + export VALGRIND=OFF
2025-12-04T09:37:14.6434912Z + VALGRIND=OFF
2025-12-04T09:37:14.6435003Z + rocminfo
2025-12-04T09:37:14.6540141Z ROCk module version 6.12.12 is loaded
2025-12-04T09:37:14.6929915Z =====================    
2025-12-04T09:37:14.6930303Z HSA System Attributes    
2025-12-04T09:37:14.6930647Z =====================    
2025-12-04T09:37:14.6930934Z Runtime Version:         1.18
2025-12-04T09:37:14.6931233Z Runtime Ext Version:     1.14
2025-12-04T09:37:14.6931566Z System Timestamp Freq.:  1000.000000MHz
2025-12-04T09:37:14.6932092Z Sig. Max Wait Duration:  18446744073709551615 (0xFFFFFFFFFFFFFFFF) (timestamp count)
2025-12-04T09:37:14.6932657Z Machine Model:           LARGE                              
2025-12-04T09:37:14.6933117Z System Endianness:       LITTLE                             
2025-12-04T09:37:14.6933512Z Mwaitx:                  DISABLED
2025-12-04T09:37:14.6933823Z XNACK enabled:           NO
2025-12-04T09:37:14.6934133Z DMAbuf Support:          YES
2025-12-04T09:37:14.6934429Z VMM Support:             YES
2025-12-04T09:37:14.6934620Z 
2025-12-04T09:37:14.6934729Z ==========               
2025-12-04T09:37:14.6935007Z HSA Agents               
2025-12-04T09:37:14.6935262Z ==========               
2025-12-04T09:37:14.6935428Z *******                  
2025-12-04T09:37:14.6935531Z Agent 1                  
2025-12-04T09:37:14.6935631Z *******                  
2025-12-04T09:37:14.6935758Z   Name:                    AMD EPYC 9575F 64-Core Processor   
2025-12-04T09:37:14.6935913Z   Uuid:                    CPU-XX                             
2025-12-04T09:37:14.6936076Z   Marketing Name:          AMD EPYC 9575F 64-Core Processor   
2025-12-04T09:37:14.6936241Z   Vendor Name:             CPU                                
2025-12-04T09:37:14.6936400Z   Feature:                 None specified                     
2025-12-04T09:37:14.6936557Z   Profile:                 FULL_PROFILE                       
2025-12-04T09:37:14.6936721Z   Float Round Mode:        NEAR                               
2025-12-04T09:37:14.6937005Z   Max Queue Number:        0(0x0)                             
2025-12-04T09:37:14.6937164Z   Queue Min Size:          0(0x0)                             
2025-12-04T09:37:14.6937317Z   Queue Max Size:          0(0x0)                             
2025-12-04T09:37:14.6937471Z   Queue Type:              MULTI                              
2025-12-04T09:37:14.6937617Z   Node:                    0                                  
2025-12-04T09:37:14.6937765Z   Device Type:             CPU                                
2025-12-04T09:37:14.6937902Z   Cache Info:              
2025-12-04T09:37:14.6938028Z     L1:                      49152(0xc000) KB                   
2025-12-04T09:37:14.6938172Z   Chip ID:                 0(0x0)                             
2025-12-04T09:37:14.6938324Z   ASIC Revision:           0(0x0)                             
2025-12-04T09:37:14.6938534Z   Cacheline Size:          64(0x40)                           
2025-12-04T09:37:14.6938696Z   Max Clock Freq. (MHz):   3300                               
2025-12-04T09:37:14.6938847Z   BDFID:                   0                                  
2025-12-04T09:37:14.6938998Z   Internal Node ID:        0                                  
2025-12-04T09:37:14.6948525Z   Compute Unit:            64                                 
2025-12-04T09:37:14.6948716Z   SIMDs per CU:            0                                  
2025-12-04T09:37:14.6948877Z   Shader Engines:          0                                  
2025-12-04T09:37:14.6949038Z   Shader Arrs. per Eng.:   0                                  
2025-12-04T09:37:14.6949203Z   WatchPts on Addr. Ranges:1                                  
2025-12-04T09:37:14.6949353Z   Memory Properties:       
2025-12-04T09:37:14.6949471Z   Features:                None
2025-12-04T09:37:14.6949586Z   Pool Info:               
2025-12-04T09:37:14.6949697Z     Pool 1                   
2025-12-04T09:37:14.6949844Z       Segment:                 GLOBAL; FLAGS: FINE GRAINED        
2025-12-04T09:37:14.6950006Z       Size:                    1584733168(0x5e751bf0) KB          
2025-12-04T09:37:14.6950162Z       Allocatable:             TRUE                               
2025-12-04T09:37:14.6950322Z       Alloc Granule:           4KB                                
2025-12-04T09:37:14.6950525Z       Alloc Recommended Granule:4KB                                
2025-12-04T09:37:14.6950694Z       Alloc Alignment:         4KB                                
2025-12-04T09:37:14.6950859Z       Accessible by all:       TRUE                               
2025-12-04T09:37:14.6951005Z     Pool 2                   
2025-12-04T09:37:14.6951141Z       Segment:                 GLOBAL; FLAGS: EXTENDED FINE GRAINED
2025-12-04T09:37:14.6951296Z       Size:                    1584733168(0x5e751bf0) KB          
2025-12-04T09:37:14.6951448Z       Allocatable:             TRUE                               
2025-12-04T09:37:14.6951608Z       Alloc Granule:           4KB                                
2025-12-04T09:37:14.6951775Z       Alloc Recommended Granule:4KB                                
2025-12-04T09:37:14.6951941Z       Alloc Alignment:         4KB                                
2025-12-04T09:37:14.6952104Z       Accessible by all:       TRUE                               
2025-12-04T09:37:14.6952244Z     Pool 3                   
2025-12-04T09:37:14.6952378Z       Segment:                 GLOBAL; FLAGS: KERNARG, FINE GRAINED
2025-12-04T09:37:14.6952530Z       Size:                    1584733168(0x5e751bf0) KB          
2025-12-04T09:37:14.6952677Z       Allocatable:             TRUE                               
2025-12-04T09:37:14.6952842Z       Alloc Granule:           4KB                                
2025-12-04T09:37:14.6953008Z       Alloc Recommended Granule:4KB                                
2025-12-04T09:37:14.6953176Z       Alloc Alignment:         4KB                                
2025-12-04T09:37:14.6953341Z       Accessible by all:       TRUE                               
2025-12-04T09:37:14.6953576Z     Pool 4                   
2025-12-04T09:37:14.6953710Z       Segment:                 GLOBAL; FLAGS: COARSE GRAINED      
2025-12-04T09:37:14.6953863Z       Size:                    1584733168(0x5e751bf0) KB          
2025-12-04T09:37:14.6954015Z       Allocatable:             TRUE                               
2025-12-04T09:37:14.6954174Z       Alloc Granule:           4KB                                
2025-12-04T09:37:14.6954338Z       Alloc Recommended Granule:4KB                                
2025-12-04T09:37:14.6954505Z       Alloc Alignment:         4KB                                
2025-12-04T09:37:14.6954666Z       Accessible by all:       TRUE                               
2025-12-04T09:37:14.6954810Z   ISA Info:                
2025-12-04T09:37:14.6954919Z *******                  
2025-12-04T09:37:14.6955026Z Agent 2                  
2025-12-04T09:37:14.6955179Z *******                  
2025-12-04T09:37:14.6955304Z   Name:                    AMD EPYC 9575F 64-Core Processor   
2025-12-04T09:37:14.6955459Z   Uuid:                    CPU-XX                             
2025-12-04T09:37:14.6955618Z   Marketing Name:          AMD EPYC 9575F 64-Core Processor   
2025-12-04T09:37:14.6955783Z   Vendor Name:             CPU                                
2025-12-04T09:37:14.6955939Z   Feature:                 None specified                     
2025-12-04T09:37:14.6956096Z   Profile:                 FULL_PROFILE                       
2025-12-04T09:37:14.6956254Z   Float Round Mode:        NEAR                               
2025-12-04T09:37:14.6956411Z   Max Queue Number:        0(0x0)                             
2025-12-04T09:37:14.6956566Z   Queue Min Size:          0(0x0)                             
2025-12-04T09:37:14.6956719Z   Queue Max Size:          0(0x0)                             
2025-12-04T09:37:14.6956874Z   Queue Type:              MULTI                              
2025-12-04T09:37:14.6957023Z   Node:                    1                                  
2025-12-04T09:37:14.6957172Z   Device Type:             CPU                                
2025-12-04T09:37:14.6957312Z   Cache Info:              
2025-12-04T09:37:14.6957435Z     L1:                      49152(0xc000) KB                   
2025-12-04T09:37:14.6957577Z   Chip ID:                 0(0x0)                             
2025-12-04T09:37:14.6957739Z   ASIC Revision:           0(0x0)                             
2025-12-04T09:37:14.6957897Z   Cacheline Size:          64(0x40)                           
2025-12-04T09:37:14.6958054Z   Max Clock Freq. (MHz):   3300                               
2025-12-04T09:37:14.6958205Z   BDFID:                   0                                  
2025-12-04T09:37:14.6958354Z   Internal Node ID:        1                                  
2025-12-04T09:37:14.6958511Z   Compute Unit:            64                                 
2025-12-04T09:37:14.6958666Z   SIMDs per CU:            0                                  
2025-12-04T09:37:14.6958822Z   Shader Engines:          0                                  
2025-12-04T09:37:14.6958984Z   Shader Arrs. per Eng.:   0                                  
2025-12-04T09:37:14.6959150Z   WatchPts on Addr. Ranges:1                                  
2025-12-04T09:37:14.6959299Z   Memory Properties:       
2025-12-04T09:37:14.6959414Z   Features:                None
2025-12-04T09:37:14.6959528Z   Pool Info:               
2025-12-04T09:37:14.6959638Z     Pool 1                   
2025-12-04T09:37:14.6959774Z       Segment:                 GLOBAL; FLAGS: FINE GRAINED        
2025-12-04T09:37:14.6959930Z       Size:                    1585355648(0x5e7e9b80) KB          
2025-12-04T09:37:14.6960083Z       Allocatable:             TRUE                               
2025-12-04T09:37:14.6960243Z       Alloc Granule:           4KB                                
2025-12-04T09:37:14.6960450Z       Alloc Recommended Granule:4KB                                
2025-12-04T09:37:14.6960618Z       Alloc Alignment:         4KB                                
2025-12-04T09:37:14.6960833Z       Accessible by all:       TRUE                               
2025-12-04T09:37:14.6960977Z     Pool 2                   
2025-12-04T09:37:14.6961115Z       Segment:                 GLOBAL; FLAGS: EXTENDED FINE GRAINED
2025-12-04T09:37:14.6961271Z       Size:                    1585355648(0x5e7e9b80) KB          
2025-12-04T09:37:14.6961422Z       Allocatable:             TRUE                               
2025-12-04T09:37:14.6961582Z       Alloc Granule:           4KB                                
2025-12-04T09:37:14.6961750Z       Alloc Recommended Granule:4KB                                
2025-12-04T09:37:14.6961915Z       Alloc Alignment:         4KB                                
2025-12-04T09:37:14.6962078Z       Accessible by all:       TRUE                               
2025-12-04T09:37:14.6962218Z     Pool 3                   
2025-12-04T09:37:14.6962428Z       Segment:                 GLOBAL; FLAGS: KERNARG, FINE GRAINED
2025-12-04T09:37:14.6962585Z       Size:                    1585355648(0x5e7e9b80) KB          
2025-12-04T09:37:14.6962733Z       Allocatable:             TRUE                               
2025-12-04T09:37:14.6962890Z       Alloc Granule:           4KB                                
2025-12-04T09:37:14.6963053Z       Alloc Recommended Granule:4KB                                
2025-12-04T09:37:14.6963215Z       Alloc Alignment:         4KB                                
2025-12-04T09:37:14.6963375Z       Accessible by all:       TRUE                               
2025-12-04T09:37:14.6963513Z     Pool 4                   
2025-12-04T09:37:14.6963642Z       Segment:                 GLOBAL; FLAGS: COARSE GRAINED      
2025-12-04T09:37:14.6963792Z       Size:                    1585355648(0x5e7e9b80) KB          
2025-12-04T09:37:14.6963940Z       Allocatable:             TRUE                               
2025-12-04T09:37:14.6964099Z       Alloc Granule:           4KB                                
2025-12-04T09:37:14.6964264Z       Alloc Recommended Granule:4KB                                
2025-12-04T09:37:14.6964425Z       Alloc Alignment:         4KB                                
2025-12-04T09:37:14.6964585Z       Accessible by all:       TRUE                               
2025-12-04T09:37:14.6964724Z   ISA Info:                
2025-12-04T09:37:14.6964828Z *******                  
2025-12-04T09:37:14.6964931Z Agent 3                  
2025-12-04T09:37:14.6965033Z *******                  
2025-12-04T09:37:14.6965148Z   Name:                    gfx942                             
2025-12-04T09:37:14.6965292Z   Uuid:                    GPU-4158e280e9a05390               
2025-12-04T09:37:14.6965447Z   Marketing Name:          AMD Radeon Graphics                
2025-12-04T09:37:14.6965606Z   Vendor Name:             AMD                                
2025-12-04T09:37:14.6965757Z   Feature:                 KERNEL_DISPATCH                    
2025-12-04T09:37:14.6965917Z   Profile:                 BASE_PROFILE                       
2025-12-04T09:37:14.6966074Z   Float Round Mode:        NEAR                               
2025-12-04T09:37:14.6966229Z   Max Queue Number:        128(0x80)                          
2025-12-04T09:37:14.6966382Z   Queue Min Size:          64(0x40)                           
2025-12-04T09:37:14.6966531Z   Queue Max Size:          131072(0x20000)                    
2025-12-04T09:37:14.6966959Z   Queue Type:              MULTI                              
2025-12-04T09:37:14.6967103Z   Node:                    2                                  
2025-12-04T09:37:14.6967246Z   Device Type:             GPU                                
2025-12-04T09:37:14.6967383Z   Cache Info:              
2025-12-04T09:37:14.6967501Z     L1:                      32(0x20) KB                        
2025-12-04T09:37:14.6967636Z     L2:                      4096(0x1000) KB                    
2025-12-04T09:37:14.6967771Z     L3:                      262144(0x40000) KB                 
2025-12-04T09:37:14.6967939Z   Chip ID:                 29861(0x74a5)                      
2025-12-04T09:37:14.6968090Z   ASIC Revision:           1(0x1)                             
2025-12-04T09:37:14.6968247Z   Cacheline Size:          128(0x80)                          
2025-12-04T09:37:14.6968402Z   Max Clock Freq. (MHz):   2100                               
2025-12-04T09:37:14.6968550Z   BDFID:                   5376                               
2025-12-04T09:37:14.6968698Z   Internal Node ID:        2                                  
2025-12-04T09:37:14.6968851Z   Compute Unit:            304                                
2025-12-04T09:37:14.6969001Z   SIMDs per CU:            4                                  
2025-12-04T09:37:14.6969152Z   Shader Engines:          32                                 
2025-12-04T09:37:14.6969307Z   Shader Arrs. per Eng.:   1                                  
2025-12-04T09:37:14.6969497Z   WatchPts on Addr. Ranges:4                                  
2025-12-04T09:37:14.6969664Z   Coherent Host Access:    FALSE                              
2025-12-04T09:37:14.6969808Z   Memory Properties:       
2025-12-04T09:37:14.6969926Z   Features:                KERNEL_DISPATCH 
2025-12-04T09:37:14.6970070Z   Fast F16 Operation:      TRUE                               
2025-12-04T09:37:14.6970229Z   Wavefront Size:          64(0x40)                           
2025-12-04T09:37:14.6970454Z   Workgroup Max Size:      1024(0x400)                        
2025-12-04T09:37:14.6970602Z   Workgroup Max Size per Dimension:
2025-12-04T09:37:14.6970732Z     x                        1024(0x400)                        
2025-12-04T09:37:14.6970864Z     y                        1024(0x400)                        
2025-12-04T09:37:14.6970993Z     z                        1024(0x400)                        
2025-12-04T09:37:14.6971137Z   Max Waves Per CU:        32(0x20)                           
2025-12-04T09:37:14.6971299Z   Max Work-item Per CU:    2048(0x800)                        
2025-12-04T09:37:14.6971461Z   Grid Max Size:           4294967295(0xffffffff)             
2025-12-04T09:37:14.6971601Z   Grid Max Size per Dimension:
2025-12-04T09:37:14.6971720Z     x                        2147483647(0x7fffffff)             
2025-12-04T09:37:14.6971854Z     y                        65535(0xffff)                      
2025-12-04T09:37:14.6971987Z     z                        65535(0xffff)                      
2025-12-04T09:37:14.6972136Z   Max fbarriers/Workgrp:   32                                 
2025-12-04T09:37:14.6972338Z   Packet Processor uCode:: 185                                
2025-12-04T09:37:14.6972502Z   SDMA engine uCode::      24                                 
2025-12-04T09:37:14.6972659Z   IOMMU Support::          None                               
2025-12-04T09:37:14.6972797Z   Pool Info:               
2025-12-04T09:37:14.6972905Z     Pool 1                   
2025-12-04T09:37:14.6973043Z       Segment:                 GLOBAL; FLAGS: COARSE GRAINED      
2025-12-04T09:37:14.6973201Z       Size:                    268419072(0xfffc000) KB            
2025-12-04T09:37:14.6973353Z       Allocatable:             TRUE                               
2025-12-04T09:37:14.6973511Z       Alloc Granule:           4KB                                
2025-12-04T09:37:14.6973676Z       Alloc Recommended Granule:2048KB                             
2025-12-04T09:37:14.6973841Z       Alloc Alignment:         4KB                                
2025-12-04T09:37:14.6974001Z       Accessible by all:       FALSE                              
2025-12-04T09:37:14.6974141Z     Pool 2                   
2025-12-04T09:37:14.6974273Z       Segment:                 GLOBAL; FLAGS: EXTENDED FINE GRAINED
2025-12-04T09:37:14.6974426Z       Size:                    268419072(0xfffc000) KB            
2025-12-04T09:37:14.6974574Z       Allocatable:             TRUE                               
2025-12-04T09:37:14.6974732Z       Alloc Granule:           4KB                                
2025-12-04T09:37:14.6974936Z       Alloc Recommended Granule:2048KB                             
2025-12-04T09:37:14.6975100Z       Alloc Alignment:         4KB                                
2025-12-04T09:37:14.6975259Z       Accessible by all:       FALSE                              
2025-12-04T09:37:14.6975399Z     Pool 3                   
2025-12-04T09:37:14.6975530Z       Segment:                 GLOBAL; FLAGS: FINE GRAINED        
2025-12-04T09:37:14.6975679Z       Size:                    268419072(0xfffc000) KB            
2025-12-04T09:37:14.6975827Z       Allocatable:             TRUE                               
2025-12-04T09:37:14.6975983Z       Alloc Granule:           4KB                                
2025-12-04T09:37:14.6976145Z       Alloc Recommended Granule:2048KB                             
2025-12-04T09:37:14.6976307Z       Alloc Alignment:         4KB                                
2025-12-04T09:37:14.6976510Z       Accessible by all:       FALSE                              
2025-12-04T09:37:14.6976649Z     Pool 4                   
2025-12-04T09:37:14.6976779Z       Segment:                 GROUP                              
2025-12-04T09:37:14.6976922Z       Size:                    64(0x40) KB                        
2025-12-04T09:37:14.6977068Z       Allocatable:             FALSE                              
2025-12-04T09:37:14.6977225Z       Alloc Granule:           0KB                                
2025-12-04T09:37:14.6977390Z       Alloc Recommended Granule:0KB                                
2025-12-04T09:37:14.6977554Z       Alloc Alignment:         0KB                                
2025-12-04T09:37:14.6977713Z       Accessible by all:       FALSE                              
2025-12-04T09:37:14.6977854Z   ISA Info:                
2025-12-04T09:37:14.6977959Z     ISA 1                    
2025-12-04T09:37:14.6978097Z       Name:                    amdgcn-amd-amdhsa--gfx942:sramecc+:xnack-
2025-12-04T09:37:14.6978269Z       Machine Models:          HSA_MACHINE_MODEL_LARGE            
2025-12-04T09:37:14.6978437Z       Profiles:                HSA_PROFILE_BASE                   
2025-12-04T09:37:14.6978600Z       Default Rounding Mode:   NEAR                               
2025-12-04T09:37:14.6978767Z       Default Rounding Mode:   NEAR                               
2025-12-04T09:37:14.6978923Z       Fast f16:                TRUE                               
2025-12-04T09:37:14.6979078Z       Workgroup Max Size:      1024(0x400)                        
2025-12-04T09:37:14.6979225Z       Workgroup Max Size per Dimension:
2025-12-04T09:37:14.6979361Z         x                        1024(0x400)                        
2025-12-04T09:37:14.6979497Z         y                        1024(0x400)                        
2025-12-04T09:37:14.6979631Z         z                        1024(0x400)                        
2025-12-04T09:37:14.6979775Z       Grid Max Size:           4294967295(0xffffffff)             
2025-12-04T09:37:14.6979920Z       Grid Max Size per Dimension:
2025-12-04T09:37:14.6980048Z         x                        2147483647(0x7fffffff)             
2025-12-04T09:37:14.6980182Z         y                        65535(0xffff)                      
2025-12-04T09:37:14.6980313Z         z                        65535(0xffff)                      
2025-12-04T09:37:14.6980497Z       FBarrier Max Size:       32                                 
2025-12-04T09:37:14.6980635Z     ISA 2                    
2025-12-04T09:37:14.6980781Z       Name:                    amdgcn-amd-amdhsa--gfx9-4-generic:sramecc+:xnack-
2025-12-04T09:37:14.6980959Z       Machine Models:          HSA_MACHINE_MODEL_LARGE            
2025-12-04T09:37:14.6981124Z       Profiles:                HSA_PROFILE_BASE                   
2025-12-04T09:37:14.6981285Z       Default Rounding Mode:   NEAR                               
2025-12-04T09:37:14.6981451Z       Default Rounding Mode:   NEAR                               
2025-12-04T09:37:14.6981608Z       Fast f16:                TRUE                               
2025-12-04T09:37:14.6981810Z       Workgroup Max Size:      1024(0x400)                        
2025-12-04T09:37:14.6981960Z       Workgroup Max Size per Dimension:
2025-12-04T09:37:14.6982090Z         x                        1024(0x400)                        
2025-12-04T09:37:14.6982223Z         y                        1024(0x400)                        
2025-12-04T09:37:14.6982354Z         z                        1024(0x400)                        
2025-12-04T09:37:14.6982499Z       Grid Max Size:           4294967295(0xffffffff)             
2025-12-04T09:37:14.6982639Z       Grid Max Size per Dimension:
2025-12-04T09:37:14.6982765Z         x                        2147483647(0x7fffffff)             
2025-12-04T09:37:14.6982897Z         y                        65535(0xffff)                      
2025-12-04T09:37:14.6983030Z         z                        65535(0xffff)                      
2025-12-04T09:37:14.6983210Z       FBarrier Max Size:       32                                 
2025-12-04T09:37:14.6983349Z *** Done ***             
2025-12-04T09:37:14.6997933Z + rocminfo
2025-12-04T09:37:14.6998346Z + grep -E 'Name:.*\sgfx|Marketing'
2025-12-04T09:37:14.7504421Z   Marketing Name:          AMD EPYC 9575F 64-Core Processor   
2025-12-04T09:37:14.7504714Z   Marketing Name:          AMD EPYC 9575F 64-Core Processor   
2025-12-04T09:37:14.7504932Z   Name:                    gfx942                             
2025-12-04T09:37:14.7505142Z   Marketing Name:          AMD Radeon Graphics                
2025-12-04T09:37:14.7545887Z + MAYBE_ROCM=rocm/
2025-12-04T09:37:14.7546070Z + [[ linux-noble-rocm-py3.12-mi300 == *xpu* ]]
2025-12-04T09:37:14.7546298Z + [[ linux-noble-rocm-py3.12-mi300 != *-bazel-* ]]
2025-12-04T09:37:14.7546485Z + pip_install ninja==1.10.2
2025-12-04T09:37:14.7546683Z + pip_install_pkg='python3 -m pip install --progress-bar off'
2025-12-04T09:37:14.7546927Z + python3 -m pip install --progress-bar off ninja==1.10.2
2025-12-04T09:37:14.9286410Z Collecting ninja==1.10.2
2025-12-04T09:37:14.9517453Z   Downloading ninja-1.10.2-py2.py3-none-manylinux_2_5_x86_64.manylinux1_x86_64.whl.metadata (5.0 kB)
2025-12-04T09:37:14.9590933Z Downloading ninja-1.10.2-py2.py3-none-manylinux_2_5_x86_64.manylinux1_x86_64.whl (108 kB)
2025-12-04T09:37:15.0544186Z Installing collected packages: ninja
2025-12-04T09:37:15.0544518Z   Attempting uninstall: ninja
2025-12-04T09:37:15.0557482Z     Found existing installation: ninja 1.11.1.4
2025-12-04T09:37:15.0566433Z     Uninstalling ninja-1.11.1.4:
2025-12-04T09:37:15.0593939Z       Successfully uninstalled ninja-1.11.1.4
2025-12-04T09:37:15.0676641Z Successfully installed ninja-1.10.2
2025-12-04T09:37:15.0984881Z + export PATH=/var/lib/jenkins/.local/bin:/opt/cache/bin:/opt/rocm/llvm/bin:/opt/rocm/opencl/bin:/opt/rocm/hip/bin:/opt/rocm/hcc/bin:/opt/rocm/bin:/opt/conda/envs/py_3.12/bin:/opt/conda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
2025-12-04T09:37:15.0986513Z + PATH=/var/lib/jenkins/.local/bin:/opt/cache/bin:/opt/rocm/llvm/bin:/opt/rocm/opencl/bin:/opt/rocm/hip/bin:/opt/rocm/hcc/bin:/opt/rocm/bin:/opt/conda/envs/py_3.12/bin:/opt/conda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
2025-12-04T09:37:15.0987317Z + [[ linux-noble-rocm-py3.12-mi300 == *aarch64* ]]
2025-12-04T09:37:15.0987610Z + [[ linux-noble-rocm-py3.12-mi300 == *asan* ]]
2025-12-04T09:37:15.0987892Z + [[ linux-noble-rocm-py3.12-mi300 == *-debug* ]]
2025-12-04T09:37:15.0988171Z + [[ linux-noble-rocm-py3.12-mi300 != *-bazel-* ]]
2025-12-04T09:37:15.0988581Z + echo 'We are not in debug mode: linux-noble-rocm-py3.12-mi300. Expect the assertion to pass'
2025-12-04T09:37:15.0989058Z We are not in debug mode: linux-noble-rocm-py3.12-mi300. Expect the assertion to pass
2025-12-04T09:37:15.0989606Z + cd test
2025-12-04T09:37:15.0989893Z + python -c 'import torch; torch._C._crash_if_debug_asserts_fail(424242)'
2025-12-04T09:37:16.3376773Z + [[ default == \n\o\g\p\u\_\N\O\_\A\V\X\2 ]]
2025-12-04T09:37:16.3377163Z + [[ default == \n\o\g\p\u\_\A\V\X\5\1\2 ]]
2025-12-04T09:37:16.3377458Z + [[ default == \l\e\g\a\c\y\_\n\v\i\d\i\a\_\d\r\i\v\e\r ]]
2025-12-04T09:37:16.3381872Z + DYNAMO_BENCHMARK_FLAGS=()
2025-12-04T09:37:16.3382137Z + [[ default == *pr_time_benchmarks* ]]
2025-12-04T09:37:16.3382387Z + [[ default == *dynamo_eager* ]]
2025-12-04T09:37:16.3382611Z + [[ default == *aot_eager* ]]
2025-12-04T09:37:16.3382824Z + [[ default == *aot_inductor* ]]
2025-12-04T09:37:16.3383060Z + [[ default == *max_autotune_inductor* ]]
2025-12-04T09:37:16.3383300Z + [[ default == *inductor* ]]
2025-12-04T09:37:16.3383516Z + [[ default == *dynamic* ]]
2025-12-04T09:37:16.3383727Z + [[ default == *cpu* ]]
2025-12-04T09:37:16.3383932Z + [[ default == *xpu* ]]
2025-12-04T09:37:16.3384179Z + DYNAMO_BENCHMARK_FLAGS+=(--device cuda)
2025-12-04T09:37:16.3400134Z + [[ linux-noble-rocm-py3.12-mi300 == *libtorch* ]]
2025-12-04T09:37:16.3400384Z + [[ linux-noble-rocm-py3.12-mi300 == *-bazel-* ]]
2025-12-04T09:37:16.3403514Z + cd test
2025-12-04T09:37:16.3404762Z + python -c 'import torch; print(torch.__config__.show())'
2025-12-04T09:37:17.0672261Z PyTorch built with:
2025-12-04T09:37:17.0672527Z   - GCC 11.5
2025-12-04T09:37:17.0672717Z   - C++ Version: 201703
2025-12-04T09:37:17.0673130Z   - Intel(R) oneAPI Math Kernel Library Version 2024.2-Product Build 20240605 for Intel(R) 64 architecture applications
2025-12-04T09:37:17.0673633Z   - Intel(R) MKL-DNN v3.7.1 (Git Hash 8d263e693366ef8db40acc569cc7d8edf644556d)
2025-12-04T09:37:17.0673956Z   - OpenMP 201511 (a.k.a. OpenMP 4.5)
2025-12-04T09:37:17.0674214Z   - LAPACK is enabled (usually provided by MKL)
2025-12-04T09:37:17.0674452Z   - NNPACK is enabled
2025-12-04T09:37:17.0674655Z   - CPU capability usage: AVX512
2025-12-04T09:37:17.0674871Z   - HIP Runtime 7.1.25424
2025-12-04T09:37:17.0675060Z   - MIOpen 3.5.1
2025-12-04T09:37:17.0675238Z   - Magma 2.9.0
2025-12-04T09:37:17.0678231Z   - Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, COMMIT_SHA=35b7a9a26c5923d98aebaa41a031dae21788a9ee, CXX_COMPILER=/opt/cache/bin/c++, CXX_FLAGS= -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -DNDEBUG -DUSE_KINETO -DLIBKINETO_NOCUPTI -DLIBKINETO_NOXPUPTI=ON -DUSE_FBGEMM -DUSE_FBGEMM_GENAI -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -O2 -fPIC -DC10_NODEPRECATED -Wall -Wextra -Werror=return-type -Werror=non-virtual-dtor -Werror=range-loop-construct -Werror=bool-operation -Wnarrowing -Wno-missing-field-initializers -Wno-unknown-pragmas -Wno-unused-parameter -Wno-strict-overflow -Wno-strict-aliasing -Wno-stringop-overflow -Wsuggest-override -Wno-psabi -Wno-error=old-style-cast -faligned-new -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Wno-stringop-overflow, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, TORCH_VERSION=2.10.0, USE_CUDA=OFF, USE_CUDNN=OFF, USE_CUSPARSELT=OFF, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_GLOO=ON, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON, USE_ROCM=ON, USE_ROCM_KERNEL_ASSERT=OFF, USE_XCCL=OFF, USE_XPU=OFF, 
2025-12-04T09:37:17.0681476Z 
2025-12-04T09:37:17.3016164Z + cd test
2025-12-04T09:37:17.3016518Z + python -c 'import torch; print(torch.__config__.parallel_info())'
2025-12-04T09:37:18.0229540Z ATen/Parallel:
2025-12-04T09:37:18.0229773Z 	at::get_num_threads() : 128
2025-12-04T09:37:18.0229934Z 	at::get_num_interop_threads() : 128
2025-12-04T09:37:18.0230069Z OpenMP 201511 (a.k.a. OpenMP 4.5)
2025-12-04T09:37:18.0230199Z 	omp_get_max_threads() : 128
2025-12-04T09:37:18.0230567Z Intel(R) oneAPI Math Kernel Library Version 2024.2-Product Build 20240605 for Intel(R) 64 architecture applications
2025-12-04T09:37:18.0230794Z 	mkl_get_max_threads() : 128
2025-12-04T09:37:18.0230959Z Intel(R) MKL-DNN v3.7.1 (Git Hash 8d263e693366ef8db40acc569cc7d8edf644556d)
2025-12-04T09:37:18.0231136Z std::thread::hardware_concurrency() : 128
2025-12-04T09:37:18.0231268Z Environment variables:
2025-12-04T09:37:18.0231383Z 	OMP_NUM_THREADS : [not set]
2025-12-04T09:37:18.0231497Z 	MKL_NUM_THREADS : [not set]
2025-12-04T09:37:18.0231648Z ATen parallel backend: OpenMP
2025-12-04T09:37:18.0231725Z 
2025-12-04T09:37:18.2706483Z + [[ default == *numpy_2* ]]
2025-12-04T09:37:18.2707111Z + [[ linux-noble-rocm-py3.12-mi300 == *aarch64* ]]
2025-12-04T09:37:18.2707271Z + [[ default == *backward* ]]
2025-12-04T09:37:18.2707409Z + [[ default == *libtorch_agnostic_targetting* ]]
2025-12-04T09:37:18.2707549Z + [[ default == *xla* ]]
2025-12-04T09:37:18.2707661Z + [[ default == *vllm* ]]
2025-12-04T09:37:18.2707778Z + [[ default == *executorch* ]]
2025-12-04T09:37:18.2707921Z + [[ default == \j\i\t\_\l\e\g\a\c\y ]]
2025-12-04T09:37:18.2708047Z + [[ default == \q\u\a\n\t\i\z\a\t\i\o\n ]]
2025-12-04T09:37:18.2708187Z + [[ linux-noble-rocm-py3.12-mi300 == *libtorch* ]]
2025-12-04T09:37:18.2708328Z + [[ default == distributed ]]
2025-12-04T09:37:18.2708444Z + [[ default == *operator_benchmark* ]]
2025-12-04T09:37:18.2708574Z + [[ default == *operator_microbenchmark* ]]
2025-12-04T09:37:18.2708708Z + [[ default == *attention_microbenchmark* ]]
2025-12-04T09:37:18.2708956Z + [[ default == *inductor_distributed* ]]
2025-12-04T09:37:18.2709083Z + [[ default == *inductor-halide* ]]
2025-12-04T09:37:18.2709208Z + [[ default == *inductor-pallas* ]]
2025-12-04T09:37:18.2709335Z + [[ default == *inductor-triton-cpu* ]]
2025-12-04T09:37:18.2709475Z + [[ default == *inductor-micro-benchmark* ]]
2025-12-04T09:37:18.2709617Z + [[ default == *aoti_cross_compile_for_windows* ]]
2025-12-04T09:37:18.2709751Z + [[ default == *huggingface* ]]
2025-12-04T09:37:18.2709863Z + [[ default == *timm* ]]
2025-12-04T09:37:18.2709969Z + [[ default == cachebench ]]
2025-12-04T09:37:18.2710082Z + [[ default == verify_cachebench ]]
2025-12-04T09:37:18.2710202Z + [[ default == *torchbench* ]]
2025-12-04T09:37:18.2710319Z + [[ default == *inductor_cpp_wrapper* ]]
2025-12-04T09:37:18.2710590Z + [[ default == *inductor_core* ]]
2025-12-04T09:37:18.2710707Z + [[ default == *inductor* ]]
2025-12-04T09:37:18.2710817Z + [[ default == *einops* ]]
2025-12-04T09:37:18.2710934Z + [[ default == *dynamo_core* ]]
2025-12-04T09:37:18.2711049Z + [[ default == *dynamo_wrapped* ]]
2025-12-04T09:37:18.2711186Z + [[ linux-noble-rocm-py3.12-mi300 == *rocm* ]]
2025-12-04T09:37:18.2711315Z + [[ -n '' ]]
2025-12-04T09:37:18.2711403Z + [[ 3 == 1 ]]
2025-12-04T09:37:18.2711490Z + [[ 3 == 2 ]]
2025-12-04T09:37:18.2711577Z + [[ 3 -gt 2 ]]
2025-12-04T09:37:18.2711672Z + install_torchvision
2025-12-04T09:37:18.2711774Z + local orig_preload
2025-12-04T09:37:18.2711870Z + local commit
2025-12-04T09:37:18.2711969Z ++ get_pinned_commit vision
2025-12-04T09:37:18.2712085Z ++ cat .github/ci_commit_pins/vision.txt
2025-12-04T09:37:18.2747660Z + commit=617079d944b0e72632311c30ae2bbdf1168b901e
2025-12-04T09:37:18.2748154Z + orig_preload=
2025-12-04T09:37:18.2748584Z + '[' -n '' ']'
2025-12-04T09:37:18.2748724Z + [[ linux-noble-rocm-py3.12-mi300 == *cuda* ]]
2025-12-04T09:37:18.2749008Z + pip_build_and_install git+https://github.com/pytorch/vision.git@617079d944b0e72632311c30ae2bbdf1168b901e dist/vision
2025-12-04T09:37:18.2749645Z + local build_target=git+https://github.com/pytorch/vision.git@617079d944b0e72632311c30ae2bbdf1168b901e
2025-12-04T09:37:18.2749869Z + local wheel_dir=dist/vision
2025-12-04T09:37:18.2749983Z + local found_whl=0
2025-12-04T09:37:18.2750088Z + for file in "${wheel_dir}"/*.whl
2025-12-04T09:37:18.2750211Z + [[ -f dist/vision/*.whl ]]
2025-12-04T09:37:18.2750319Z + '[' 0 == 0 ']'
2025-12-04T09:37:18.2750606Z + python3 -m pip wheel --no-build-isolation --no-deps -w dist/vision git+https://github.com/pytorch/vision.git@617079d944b0e72632311c30ae2bbdf1168b901e
2025-12-04T09:37:18.4277741Z Collecting git+https://github.com/pytorch/vision.git@617079d944b0e72632311c30ae2bbdf1168b901e
2025-12-04T09:37:18.4278464Z   Cloning https://github.com/pytorch/vision.git (to revision 617079d944b0e72632311c30ae2bbdf1168b901e) to /tmp/pip-req-build-1itotbq1
2025-12-04T09:37:18.4304797Z   Running command git clone --filter=blob:none --quiet https://github.com/pytorch/vision.git /tmp/pip-req-build-1itotbq1
2025-12-04T09:37:21.6893563Z   Running command git rev-parse -q --verify 'sha^617079d944b0e72632311c30ae2bbdf1168b901e'
2025-12-04T09:37:21.6911186Z   Running command git fetch -q https://github.com/pytorch/vision.git 617079d944b0e72632311c30ae2bbdf1168b901e
2025-12-04T09:37:22.4051493Z   Resolved https://github.com/pytorch/vision.git to commit 617079d944b0e72632311c30ae2bbdf1168b901e
2025-12-04T09:37:23.9355677Z   Preparing metadata (pyproject.toml) ... [?25l- \ | done
2025-12-04T09:37:23.9373908Z [?25hBuilding wheels for collected packages: torchvision
2025-12-04T09:38:03.2069575Z   Building wheel for torchvision (pyproject.toml) ... [?25l- \ | / - \ | / - \ | / - \ | / - \ | / - \ | / - \ done
2025-12-04T09:38:03.2090343Z [?25h  Created wheel for torchvision: filename=torchvision-0.25.0a0+617079d-cp312-cp312-linux_x86_64.whl size=1814543 sha256=46315f3798682880c001cd9eac433fa27b0e38eb93c918062f77a7dbc7c07896
2025-12-04T09:38:03.2093209Z   Stored in directory: /var/lib/jenkins/.cache/pip/wheels/22/df/b5/2cdf6bb6a10c31c47b56cf4d0441cf0ee834f1c9dee15fb9d9
2025-12-04T09:38:03.2115120Z Successfully built torchvision
2025-12-04T09:38:03.2612036Z + for file in "${wheel_dir}"/*.whl
2025-12-04T09:38:03.2612421Z + pip_install_whl dist/vision/torchvision-0.25.0a0+617079d-cp312-cp312-linux_x86_64.whl
2025-12-04T09:38:03.2612864Z + args=('dist/vision/torchvision-0.25.0a0+617079d-cp312-cp312-linux_x86_64.whl')
2025-12-04T09:38:03.2613166Z + local args
2025-12-04T09:38:03.2613436Z + [[ dist/vision/torchvision-0.25.0a0+617079d-cp312-cp312-linux_x86_64.whl == *\ * ]]
2025-12-04T09:38:03.2613768Z + for path in "${args[@]}"
2025-12-04T09:38:03.2614085Z + echo 'Installing dist/vision/torchvision-0.25.0a0+617079d-cp312-cp312-linux_x86_64.whl'
2025-12-04T09:38:03.2614510Z Installing dist/vision/torchvision-0.25.0a0+617079d-cp312-cp312-linux_x86_64.whl
2025-12-04T09:38:03.2615014Z + python3 -mpip install --no-index --no-deps dist/vision/torchvision-0.25.0a0+617079d-cp312-cp312-linux_x86_64.whl
2025-12-04T09:38:03.3996609Z Processing ./dist/vision/torchvision-0.25.0a0+617079d-cp312-cp312-linux_x86_64.whl
2025-12-04T09:38:03.4039949Z Installing collected packages: torchvision
2025-12-04T09:38:03.6338167Z Successfully installed torchvision-0.25.0a0+617079d
2025-12-04T09:38:03.6598404Z + '[' -n '' ']'
2025-12-04T09:38:03.6599022Z + test_python_shard 3
2025-12-04T09:38:03.6599376Z + [[ -z 6 ]]
2025-12-04T09:38:03.6600194Z + python test/run_test.py --exclude-jit-executor --exclude-distributed-tests --exclude-quantization-tests --shard 3 6 --verbose --upload-artifacts-while-running
2025-12-04T09:38:05.2609636Z Excluding inductor/test_max_autotune on ROCm
2025-12-04T09:38:05.2610108Z Excluding test_cuda_nvml_based_avail on ROCm
2025-12-04T09:38:05.6115492Z Downloading https://ossci-metrics.s3.amazonaws.com/disabled-tests-condensed.json to /var/lib/jenkins/pytorch/test/.pytorch-disabled-tests.json
2025-12-04T09:38:05.9742568Z Ignoring disabled issues:  ['']
2025-12-04T09:38:05.9742802Z Found test times from artifacts
2025-12-04T09:38:05.9921242Z Found test times from artifacts
2025-12-04T09:38:05.9921438Z Running all tests
2025-12-04T09:38:06.0072301Z Running parallel tests on 1 processes
2025-12-04T09:38:06.0083154Z Name: tests to run (est. time: 174.68min)
2025-12-04T09:38:06.0083325Z   Serial tests (102):
2025-12-04T09:38:06.0083441Z     inductor/test_aot_inductor 2/3
2025-12-04T09:38:06.0083581Z     inductor/test_torchinductor_dynamic_shapes 5/5
2025-12-04T09:38:06.0083732Z     inductor/test_torchinductor_opinfo 3/10
2025-12-04T09:38:06.0083881Z     inductor/test_torchinductor_opinfo 9/10
2025-12-04T09:38:06.0084010Z     inductor/test_cpu_repro 3/4
2025-12-04T09:38:06.0084128Z     dynamo/test_higher_order_ops 1/1
2025-12-04T09:38:06.0084255Z     inductor/test_custom_lowering 1/1
2025-12-04T09:38:06.0084377Z     inductor/test_fused_attention 1/1
2025-12-04T09:38:06.0084496Z     inductor/test_smoke 1/1
2025-12-04T09:38:06.0084608Z     inductor/test_flex_attention 1/4
2025-12-04T09:38:06.0084729Z     inductor/test_cutlass_backend 1/1
2025-12-04T09:38:06.0084850Z     inductor/test_custom_op_autotune 1/1
2025-12-04T09:38:06.0085327Z     inductor/test_compile_subprocess 2/3
2025-12-04T09:38:06.0085457Z     dynamo/test_model_output 1/1
2025-12-04T09:38:06.0085576Z     inductor/test_selective_lowering 1/1
2025-12-04T09:38:06.0085699Z     dynamo/test_backends 1/1
2025-12-04T09:38:06.0085815Z     inductor/test_triton_heuristics 1/1
2025-12-04T09:38:06.0085944Z     inductor/test_flex_decoding 2/2
2025-12-04T09:38:06.0086064Z     inductor/test_b2b_gemm 1/1
2025-12-04T09:38:06.0086178Z     export/test_unflatten 1/1
2025-12-04T09:38:06.0086289Z     export/test_hop 1/1
2025-12-04T09:38:06.0086394Z     export/test_serdes 1/1
2025-12-04T09:38:06.0086507Z     inductor/test_debug_trace 1/1
2025-12-04T09:38:06.0086629Z     dynamo/test_guard_serialization 1/1
2025-12-04T09:38:06.0086752Z     dynamo/test_recompile_ux 1/1
2025-12-04T09:38:06.0086865Z     export/test_torchbind 1/1
2025-12-04T09:38:06.0087095Z     export/test_strict_export_v2 1/1
2025-12-04T09:38:06.0087218Z     dynamo/test_structured_trace 1/1
2025-12-04T09:38:06.0087339Z     export/test_export_strict 1/1
2025-12-04T09:38:06.0087458Z     dynamo/test_autograd_function 1/1
2025-12-04T09:38:06.0087588Z     dynamo/test_backward_higher_order_ops 1/1
2025-12-04T09:38:06.0087717Z     dynamo/test_base_hop 1/1
2025-12-04T09:38:06.0087828Z     dynamo/test_base_output 1/1
2025-12-04T09:38:06.0087943Z     dynamo/test_buffers_override 1/1
2025-12-04T09:38:06.0088062Z     dynamo/test_bytecode_utils 1/1
2025-12-04T09:38:06.0088176Z     dynamo/test_callback 1/1
2025-12-04T09:38:06.0088284Z     dynamo/test_compile 1/1
2025-12-04T09:38:06.0088395Z     dynamo/test_compiler_bisector 1/1
2025-12-04T09:38:06.0088514Z     dynamo/test_comptime 1/1
2025-12-04T09:38:06.0088620Z     dynamo/test_config 1/1
2025-12-04T09:38:06.0088727Z     dynamo/test_debug_utils 1/1
2025-12-04T09:38:06.0088840Z     dynamo/test_deque_reconstruct 1/1
2025-12-04T09:38:06.0088958Z     dynamo/test_deviceguard 1/1
2025-12-04T09:38:06.0089070Z     dynamo/test_dicts 1/1
2025-12-04T09:38:06.0089176Z     dynamo/test_exceptions 1/1
2025-12-04T09:38:06.0089291Z     dynamo/test_export_mutations 1/1
2025-12-04T09:38:06.0089410Z     dynamo/test_flat_apply 1/1
2025-12-04T09:38:06.0089520Z     dynamo/test_frame_init 1/1
2025-12-04T09:38:06.0089631Z     dynamo/test_fx_annotate 1/1
2025-12-04T09:38:06.0089741Z     dynamo/test_interop 1/1
2025-12-04T09:38:06.0089852Z     dynamo/test_list 1/1
2025-12-04T09:38:06.0089960Z     dynamo/test_metrics_context 1/1
2025-12-04T09:38:06.0090076Z     dynamo/test_minifier 1/1
2025-12-04T09:38:06.0090185Z     dynamo/test_modes 1/1
2025-12-04T09:38:06.0090291Z     dynamo/test_optimizers 1/1
2025-12-04T09:38:06.0090532Z     dynamo/test_pre_dispatch 1/1
2025-12-04T09:38:06.0090650Z     dynamo/test_precompile_context 1/1
2025-12-04T09:38:06.0090769Z     dynamo/test_profiler 1/1
2025-12-04T09:38:06.0090878Z     dynamo/test_reorder_logs 1/1
2025-12-04T09:38:06.0090987Z     dynamo/test_sets 1/1
2025-12-04T09:38:06.0091102Z     dynamo/test_skip_guard_eval_unsafe 1/1
2025-12-04T09:38:06.0091224Z     dynamo/test_trace_rules 1/1
2025-12-04T09:38:06.0091337Z     dynamo/test_tree_map 1/1
2025-12-04T09:38:06.0091460Z     dynamo/test_wrap_inductor_compiled_regions 1/1
2025-12-04T09:38:06.0091617Z     export/test_upgrader 1/1
2025-12-04T09:38:06.0091732Z     export/test_verifier 1/1
2025-12-04T09:38:06.0091848Z     inductor/test_alignment 1/1
2025-12-04T09:38:06.0091968Z     inductor/test_best_config 1/1
2025-12-04T09:38:06.0092093Z     inductor/test_block_analysis 1/1
2025-12-04T09:38:06.0092216Z     inductor/test_cache 1/1
2025-12-04T09:38:06.0092330Z     inductor/test_compile 1/1
2025-12-04T09:38:06.0092450Z     inductor/test_compile_worker 1/1
2025-12-04T09:38:06.0092575Z     inductor/test_control_flow 2/4
2025-12-04T09:38:06.0092696Z     inductor/test_pallas 1/1
2025-12-04T09:38:06.0092810Z     test_varlen_attention 1/1
2025-12-04T09:38:06.0092922Z     test_torch 1/1
2025-12-04T09:38:06.0093028Z     test_utils_filelock 1/1
2025-12-04T09:38:06.0093141Z     test_fake_tensor 1/1
2025-12-04T09:38:06.0093245Z     test_ops 1/5
2025-12-04T09:38:06.0093385Z     test_decomp 2/11
2025-12-04T09:38:06.0093488Z     test_decomp 8/11
2025-12-04T09:38:06.0093589Z     test_nn 3/3
2025-12-04T09:38:06.0093744Z     cpp_extensions/libtorch_agnostic_2_10_extension/test_version_compatibility 1/1
2025-12-04T09:38:06.0093969Z     cpp_extensions/python_agnostic_extension/test/test_python_agnostic 1/1
2025-12-04T09:38:06.0094152Z     cpp_extensions/test_libtorch_agnostic 1/1
2025-12-04T09:38:06.0094292Z     distributions/test_constraints 1/1
2025-12-04T09:38:06.0094423Z     functorch/dim/test_getsetitem 1/1
2025-12-04T09:38:06.0094546Z     functorch/dim/test_split 1/1
2025-12-04T09:38:06.0094665Z     functorch/test_ac_knapsack 1/1
2025-12-04T09:38:06.0094790Z     functorch/test_ac_logging 1/1
2025-12-04T09:38:06.0094914Z     functorch/test_control_flow 1/2
2025-12-04T09:38:06.0095038Z     functorch/test_ops 5/5
2025-12-04T09:38:06.0095185Z     optim/test_optim 1/1
2025-12-04T09:38:06.0095294Z     test_jiterator 1/1
2025-12-04T09:38:06.0095400Z     test_legacy_vmap 1/1
2025-12-04T09:38:06.0095510Z     test_modules 2/2
2025-12-04T09:38:06.0095613Z     test_optim 1/1
2025-12-04T09:38:06.0095725Z     test_set_default_mobile_cpu_allocator 1/1
2025-12-04T09:38:06.0095853Z     test_shape_ops 1/1
2025-12-04T09:38:06.0095955Z     test_show_pickle 1/1
2025-12-04T09:38:06.0096063Z     test_sparse_csr 1/2
2025-12-04T09:38:06.0096167Z     xpu/test_fusion 1/1
2025-12-04T09:38:06.0096272Z   Parallel tests (0):
2025-12-04T09:38:06.0096383Z Name: excluded (est. time: 0.0min)
2025-12-04T09:38:06.0096505Z   Serial tests (0):
2025-12-04T09:38:06.0096607Z   Parallel tests (0):
2025-12-04T09:38:06.0096777Z Running inductor/test_aot_inductor 2/3 ... [2025-12-04 09:38:06.007664][2246303.141618741]
2025-12-04T09:38:06.0096966Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:38:06.0112179Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_aot_inductor.py', '--shard-id=2', '--num-shards=3', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:38:06.007865]
2025-12-04T09:38:14.1445613Z 
2025-12-04T09:38:14.1446456Z inductor/test_aot_inductor 2/3 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_aot_inductor_2.3_77c63ae7486131ff_.log
2025-12-04T09:38:14.1446957Z Running 0 items in this shard:
2025-12-04T09:38:14.1447079Z 
2025-12-04T09:38:14.1447259Z Finished inductor/test_aot_inductor 2/3 ... [2025-12-04 09:38:14.144373][2246311.278326042], took 0.14min
2025-12-04T09:38:14.1451456Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:38:16.3051316Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:38:16.3052046Z GITHUB_RUN_ID, GITHUB_RUN_ATTEMPT, or ARTIFACTS_FILE_SUFFIX not set, not uploading
2025-12-04T09:38:16.3052554Z Uploading artifacts took 0.00 seconds
2025-12-04T09:38:16.3053145Z Running inductor/test_torchinductor_dynamic_shapes 5/5 ... [2025-12-04 09:38:16.304885][2246313.438839396]
2025-12-04T09:38:16.3053732Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:38:16.3055094Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_torchinductor_dynamic_shapes.py', '--shard-id=5', '--num-shards=5', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:38:16.305238]
2025-12-04T09:39:16.6582371Z 
2025-12-04T09:39:16.6583340Z PRINTING LOG FILE of inductor/test_torchinductor_dynamic_shapes 5/5 (test/test-reports/inductor.test_torchinductor_dynamic_shapes_5.5_7bd540a7dc87d591_.log)
2025-12-04T09:39:16.6584339Z Test results will be stored in test-reports/python-pytest/inductor.test_torchinductor_dynamic_shapes/inductor.test_torchinductor_dynamic_shapes-1e4909ab018e7c6d.xml
2025-12-04T09:39:16.6585653Z ============================= test session starts ==============================
2025-12-04T09:39:16.6586111Z platform linux -- Python 3.12.5, pytest-7.3.2, pluggy-1.6.0 -- /opt/conda/envs/py_3.12/bin/python
2025-12-04T09:39:16.6586491Z cachedir: .pytest_cache
2025-12-04T09:39:16.6586929Z hypothesis profile 'pytorch_ci' -> database=None, max_examples=50, derandomize=True, suppress_health_check=[HealthCheck.too_slow]
2025-12-04T09:39:16.6587400Z rootdir: /var/lib/jenkins/pytorch
2025-12-04T09:39:16.6587618Z configfile: pytest.ini
2025-12-04T09:39:16.6588035Z plugins: hypothesis-6.56.4, cpp-2.3.0, flakefinder-1.1.0, rerunfailures-14.0, subtests-0.13.1, xdist-3.3.1, xdoctest-1.3.0, typeguard-4.3.0
2025-12-04T09:39:16.6588475Z collecting ... collected 1851 items
2025-12-04T09:39:16.6588733Z stepcurrent: Cannot find last run test, not skipping
2025-12-04T09:39:16.6603014Z Running 50 items in this shard: test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda, test/inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6613136Z 
2025-12-04T09:39:16.6613406Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [2.1502s] [  2%]
2025-12-04T09:39:16.6613940Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.1999s] [  2%]
2025-12-04T09:39:16.6614475Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.1846s] [  2%]
2025-12-04T09:39:16.6615049Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.2104s] [  2%]
2025-12-04T09:39:16.6615576Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.1909s] [  2%]
2025-12-04T09:39:16.6616103Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.1797s] [  2%]
2025-12-04T09:39:16.6616632Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.9008s] [  2%]
2025-12-04T09:39:16.6617154Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.2145s] [  2%]
2025-12-04T09:39:16.6617690Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.1796s] [  2%]
2025-12-04T09:39:16.6618195Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [1.1713s] [  2%]
2025-12-04T09:39:16.6618698Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0337s] [  2%]
2025-12-04T09:39:16.6619205Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.5407s] [  2%]
2025-12-04T09:39:16.6619707Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.5404s] [  2%]
2025-12-04T09:39:16.6620211Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0543s] [  2%]
2025-12-04T09:39:16.6620744Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.5656s] [  2%]
2025-12-04T09:39:16.6621247Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.5276s] [  2%]
2025-12-04T09:39:16.6621755Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0629s] [  2%]
2025-12-04T09:39:16.6622262Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0845s] [  2%]
2025-12-04T09:39:16.6622765Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0729s] [  2%]
2025-12-04T09:39:16.6623271Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [1.4937s] [  2%]
2025-12-04T09:39:16.6623773Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0973s] [  2%]
2025-12-04T09:39:16.6624279Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.5544s] [  2%]
2025-12-04T09:39:16.6624889Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.5585s] [  2%]
2025-12-04T09:39:16.6625399Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0659s] [  2%]
2025-12-04T09:39:16.6625907Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0621s] [  2%]
2025-12-04T09:39:16.6626439Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.5515s] [  2%]
2025-12-04T09:39:16.6626947Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0654s] [  2%]
2025-12-04T09:39:16.6627511Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0676s] [  2%]
2025-12-04T09:39:16.6628024Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0645s] [  2%]
2025-12-04T09:39:16.6628530Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.5407s] [  2%]
2025-12-04T09:39:16.6629036Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.5468s] [  2%]
2025-12-04T09:39:16.6629542Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [1.0617s] [  2%]
2025-12-04T09:39:16.6630057Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0477s] [  2%]
2025-12-04T09:39:16.6630603Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.5325s] [  2%]
2025-12-04T09:39:16.6642170Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0503s] [  2%]
2025-12-04T09:39:16.6642737Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.5359s] [  2%]
2025-12-04T09:39:16.6643278Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0600s] [  2%]
2025-12-04T09:39:16.6643795Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [1.0560s] [  2%]
2025-12-04T09:39:16.6644317Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.2317s] [  2%]
2025-12-04T09:39:16.6644841Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.2983s] [  2%]
2025-12-04T09:39:16.6645363Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.2991s] [  2%]
2025-12-04T09:39:16.6645953Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.7166s] [  2%]
2025-12-04T09:39:16.6646473Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.3473s] [  2%]
2025-12-04T09:39:16.6646991Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.7142s] [  2%]
2025-12-04T09:39:16.6647511Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [1.3510s] [  2%]
2025-12-04T09:39:16.6648030Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.6883s] [  2%]
2025-12-04T09:39:16.6648588Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py FAILED [0.6755s] [  2%]
2025-12-04T09:39:16.6649104Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.8453s] [  2%]
2025-12-04T09:39:16.6649621Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.2148s] [  2%]
2025-12-04T09:39:16.6650135Z inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda <- test/inductor/test_torchinductor.py PASSED [1.0808s] [  2%]
2025-12-04T09:39:16.6650453Z 
2025-12-04T09:39:16.6650529Z =================================== FAILURES ===================================
2025-12-04T09:39:16.6650743Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6650943Z Traceback (most recent call last):
2025-12-04T09:39:16.6651173Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6651392Z     self.common(
2025-12-04T09:39:16.6651564Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6651750Z     return func(*args, **kwds)
2025-12-04T09:39:16.6651872Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6652082Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.6652295Z     check_model(
2025-12-04T09:39:16.6652484Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6652689Z     assert_equal_fn(
2025-12-04T09:39:16.6652907Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6653163Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6653317Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6653577Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6653864Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6654042Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6654402Z 
2025-12-04T09:39:16.6654459Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6654655Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6654909Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.6655053Z 
2025-12-04T09:39:16.6655110Z The failure occurred for item [2]
2025-12-04T09:39:16.6655192Z 
2025-12-04T09:39:16.6655314Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6655654Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6655908Z 
2025-12-04T09:39:16.6656008Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6656226Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6656413Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6656844Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6657312Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6657531Z graph_break []
2025-12-04T09:39:16.6657703Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6657899Z Traceback (most recent call last):
2025-12-04T09:39:16.6658117Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6658331Z     self.common(
2025-12-04T09:39:16.6658491Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6658670Z     return func(*args, **kwds)
2025-12-04T09:39:16.6658791Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6658995Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.6659205Z     check_model(
2025-12-04T09:39:16.6659388Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6659594Z     assert_equal_fn(
2025-12-04T09:39:16.6659815Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6660067Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6660217Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6660519Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6660804Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6660977Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6661070Z 
2025-12-04T09:39:16.6661124Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6661316Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6661568Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.6661710Z 
2025-12-04T09:39:16.6661770Z The failure occurred for item [2]
2025-12-04T09:39:16.6661849Z 
2025-12-04T09:39:16.6661933Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6662263Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6662520Z 
2025-12-04T09:39:16.6662611Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6662822Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6663003Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6663430Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6663891Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6664079Z graph_break []
2025-12-04T09:39:16.6664260Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6664444Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6664649Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6665152Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6665601Z graph_break []
2025-12-04T09:39:16.6665740Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6665921Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6666159Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6666664Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6667101Z graph_break []
2025-12-04T09:39:16.6667238Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6667416Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6667616Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6668096Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6668524Z graph_break []
2025-12-04T09:39:16.6668656Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6668830Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6669026Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6669516Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6669948Z graph_break []
2025-12-04T09:39:16.6670080Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6670255Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6670485Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6670985Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6671413Z graph_break []
2025-12-04T09:39:16.6671547Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6671723Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6671919Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6672406Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6672832Z graph_break []
2025-12-04T09:39:16.6673028Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6673220Z Traceback (most recent call last):
2025-12-04T09:39:16.6673435Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6673643Z     self.common(
2025-12-04T09:39:16.6673796Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6673971Z     return func(*args, **kwds)
2025-12-04T09:39:16.6674089Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6674292Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 725, in check_model_gpu
2025-12-04T09:39:16.6674500Z     check_model(
2025-12-04T09:39:16.6674680Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6674881Z     assert_equal_fn(
2025-12-04T09:39:16.6675143Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6675385Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6675530Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6675779Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6676057Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6676224Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6676319Z 
2025-12-04T09:39:16.6676367Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6676549Z Greatest absolute difference: 0.58544921875 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6676781Z Greatest relative difference: 0.57568359375 at index (0, 1) (up to 0.001 allowed)
2025-12-04T09:39:16.6676918Z 
2025-12-04T09:39:16.6676971Z The failure occurred for item [2]
2025-12-04T09:39:16.6677056Z 
2025-12-04T09:39:16.6677135Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6677462Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6677716Z 
2025-12-04T09:39:16.6677806Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6678011Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6678187Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6678608Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6679060Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6679243Z graph_break []
2025-12-04T09:39:16.6679381Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6679560Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6679758Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6680255Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6680711Z graph_break []
2025-12-04T09:39:16.6680846Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6681023Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6681219Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6681747Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6682173Z graph_break []
2025-12-04T09:39:16.6682308Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6682484Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6682680Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6683175Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6683633Z graph_break []
2025-12-04T09:39:16.6683771Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6683948Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6684146Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6684641Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6685068Z graph_break []
2025-12-04T09:39:16.6685202Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6685377Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6685572Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6686066Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6686489Z graph_break []
2025-12-04T09:39:16.6686622Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6686795Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6686988Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6687474Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6687900Z graph_break []
2025-12-04T09:39:16.6688035Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6688210Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6688405Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6688897Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6689324Z graph_break []
2025-12-04T09:39:16.6689457Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6689631Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6689829Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6690348Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6690814Z graph_break []
2025-12-04T09:39:16.6690949Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6691125Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6691318Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6691808Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6692269Z graph_break []
2025-12-04T09:39:16.6692433Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6692623Z Traceback (most recent call last):
2025-12-04T09:39:16.6692839Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6693204Z     self.common(
2025-12-04T09:39:16.6693357Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6693530Z     return func(*args, **kwds)
2025-12-04T09:39:16.6693648Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6693850Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.6694059Z     check_model(
2025-12-04T09:39:16.6694240Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6694442Z     assert_equal_fn(
2025-12-04T09:39:16.6694657Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6694904Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6695051Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6695302Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6695578Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6695743Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6695836Z 
2025-12-04T09:39:16.6695883Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6696070Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6696315Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.6696461Z 
2025-12-04T09:39:16.6696513Z The failure occurred for item [2]
2025-12-04T09:39:16.6696596Z 
2025-12-04T09:39:16.6696672Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6696998Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6697252Z 
2025-12-04T09:39:16.6697340Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6697544Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6697717Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6698135Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6698591Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6698772Z graph_break []
2025-12-04T09:39:16.6698947Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6699125Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6699321Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6699809Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6700240Z graph_break []
2025-12-04T09:39:16.6700375Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6700595Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6700792Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6701320Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6701749Z graph_break []
2025-12-04T09:39:16.6701882Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6702056Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6702252Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6702744Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6703175Z graph_break []
2025-12-04T09:39:16.6703314Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6703487Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6703685Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6704175Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6704607Z graph_break []
2025-12-04T09:39:16.6704741Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6704915Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6705112Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6705609Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6706034Z graph_break []
2025-12-04T09:39:16.6706167Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6706340Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6706535Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6707023Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6707452Z graph_break []
2025-12-04T09:39:16.6707622Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6707795Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6707990Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6708479Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6708907Z graph_break []
2025-12-04T09:39:16.6709041Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6709215Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6709410Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6709928Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6710351Z graph_break []
2025-12-04T09:39:16.6710519Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6710694Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6710890Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6711377Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6711810Z graph_break []
2025-12-04T09:39:16.6711947Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6712123Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6712318Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6712808Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6713234Z graph_break []
2025-12-04T09:39:16.6713369Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6713546Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6713744Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6714237Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6714663Z graph_break []
2025-12-04T09:39:16.6714823Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6715012Z Traceback (most recent call last):
2025-12-04T09:39:16.6715230Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6715438Z     self.common(
2025-12-04T09:39:16.6715590Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6715764Z     return func(*args, **kwds)
2025-12-04T09:39:16.6715876Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6716080Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.6716327Z     check_model(
2025-12-04T09:39:16.6716509Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6716708Z     assert_equal_fn(
2025-12-04T09:39:16.6716914Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6717158Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6717302Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6717551Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6717831Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6717997Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6718090Z 
2025-12-04T09:39:16.6718175Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6718361Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6718606Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.6718751Z 
2025-12-04T09:39:16.6718798Z The failure occurred for item [2]
2025-12-04T09:39:16.6718881Z 
2025-12-04T09:39:16.6718956Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6719279Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6719533Z 
2025-12-04T09:39:16.6719626Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6719831Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6720005Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6720480Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6720930Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6721104Z graph_break []
2025-12-04T09:39:16.6721233Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6721405Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6721599Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6722083Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6722509Z graph_break []
2025-12-04T09:39:16.6722637Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6722807Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6722999Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6723484Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6723907Z graph_break []
2025-12-04T09:39:16.6724033Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6724204Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6724394Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6724914Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6725335Z graph_break []
2025-12-04T09:39:16.6725482Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6725651Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6725842Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6726325Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6726777Z graph_break []
2025-12-04T09:39:16.6726907Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6727075Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6727266Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6727747Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6728170Z graph_break []
2025-12-04T09:39:16.6728299Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6728466Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6728657Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6729143Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6729561Z graph_break []
2025-12-04T09:39:16.6729688Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6729858Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6730050Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6730571Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6730995Z graph_break []
2025-12-04T09:39:16.6731124Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6731296Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6731488Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6731970Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6732394Z graph_break []
2025-12-04T09:39:16.6732524Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6732693Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6732882Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6733455Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6733876Z graph_break []
2025-12-04T09:39:16.6734005Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6734174Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6734367Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6734850Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6735301Z graph_break []
2025-12-04T09:39:16.6735432Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6735602Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6735794Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6736276Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6736693Z graph_break []
2025-12-04T09:39:16.6736822Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6736990Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6737181Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6737666Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6738097Z graph_break []
2025-12-04T09:39:16.6738251Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6738436Z Traceback (most recent call last):
2025-12-04T09:39:16.6738642Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6738840Z     self.common(
2025-12-04T09:39:16.6738986Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6739154Z     return func(*args, **kwds)
2025-12-04T09:39:16.6739267Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6739467Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.6739668Z     check_model(
2025-12-04T09:39:16.6739846Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6740040Z     assert_equal_fn(
2025-12-04T09:39:16.6740241Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6740513Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6740652Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6740896Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6741165Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6741323Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6741414Z 
2025-12-04T09:39:16.6741459Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6741641Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6741924Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.6742067Z 
2025-12-04T09:39:16.6742115Z The failure occurred for item [2]
2025-12-04T09:39:16.6742194Z 
2025-12-04T09:39:16.6742268Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6742587Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6742835Z 
2025-12-04T09:39:16.6742927Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6743131Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6743299Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6743748Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6744197Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6744375Z graph_break []
2025-12-04T09:39:16.6744504Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6744674Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6744866Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6745350Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6745779Z graph_break []
2025-12-04T09:39:16.6745908Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6746080Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6746271Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6746756Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6747183Z graph_break []
2025-12-04T09:39:16.6747310Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6747481Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6747671Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6748164Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6748587Z graph_break []
2025-12-04T09:39:16.6748714Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6748884Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6749078Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6749573Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6749997Z graph_break []
2025-12-04T09:39:16.6750150Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6750320Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6750548Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6751037Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6751462Z graph_break []
2025-12-04T09:39:16.6751595Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6751772Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6751968Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6752495Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6752916Z graph_break []
2025-12-04T09:39:16.6753048Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6753221Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6753416Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6753899Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6754320Z graph_break []
2025-12-04T09:39:16.6754454Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6754626Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6754818Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6755303Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6755729Z graph_break []
2025-12-04T09:39:16.6755852Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6756021Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6756211Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6756704Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6757129Z graph_break []
2025-12-04T09:39:16.6757257Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6757423Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6757615Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6758096Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6758520Z graph_break []
2025-12-04T09:39:16.6758678Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6758846Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6759036Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6759520Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6759939Z graph_break []
2025-12-04T09:39:16.6760066Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6760235Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6760461Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6760983Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6761402Z graph_break []
2025-12-04T09:39:16.6761530Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6761696Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6761887Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6762373Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6762802Z graph_break []
2025-12-04T09:39:16.6762933Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6763101Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6763292Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6763773Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6764193Z graph_break []
2025-12-04T09:39:16.6764342Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6764525Z Traceback (most recent call last):
2025-12-04T09:39:16.6764733Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6764938Z     self.common(
2025-12-04T09:39:16.6765084Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6765256Z     return func(*args, **kwds)
2025-12-04T09:39:16.6765365Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6765561Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.6765765Z     check_model(
2025-12-04T09:39:16.6765941Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6766136Z     assert_equal_fn(
2025-12-04T09:39:16.6766337Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6766577Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6766719Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6766967Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6767279Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6767442Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6767530Z 
2025-12-04T09:39:16.6767578Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6767759Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6768001Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.6768144Z 
2025-12-04T09:39:16.6768201Z The failure occurred for item [2]
2025-12-04T09:39:16.6768281Z 
2025-12-04T09:39:16.6768356Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6768676Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6768929Z 
2025-12-04T09:39:16.6769055Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6769260Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6769429Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6769843Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6770289Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6770506Z graph_break []
2025-12-04T09:39:16.6770633Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6770804Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6770994Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6771492Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6771914Z graph_break []
2025-12-04T09:39:16.6772040Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6772208Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6772399Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6772886Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6773311Z graph_break []
2025-12-04T09:39:16.6773442Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6773611Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6773799Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6774283Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6774703Z graph_break []
2025-12-04T09:39:16.6774827Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6774994Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6775182Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6775704Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6776124Z graph_break []
2025-12-04T09:39:16.6776249Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6776416Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6776605Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6777091Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6777551Z graph_break []
2025-12-04T09:39:16.6777686Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6777862Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6778059Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6778554Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6778980Z graph_break []
2025-12-04T09:39:16.6779115Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6779290Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6779485Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6779979Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6780444Z graph_break []
2025-12-04T09:39:16.6780578Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6780751Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6780947Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6781438Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6781869Z graph_break []
2025-12-04T09:39:16.6782004Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6782179Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6782374Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6782865Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6783295Z graph_break []
2025-12-04T09:39:16.6783430Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6783605Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6783801Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6784330Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6784757Z graph_break []
2025-12-04T09:39:16.6784891Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6785066Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6785263Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6785754Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6786213Z graph_break []
2025-12-04T09:39:16.6786349Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6786523Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6786720Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6787204Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6787630Z graph_break []
2025-12-04T09:39:16.6787762Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6787937Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6788127Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6788631Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6789058Z graph_break []
2025-12-04T09:39:16.6789191Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6789364Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6789562Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6790059Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6790537Z graph_break []
2025-12-04T09:39:16.6790672Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6790851Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6791048Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6791542Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6791970Z graph_break []
2025-12-04T09:39:16.6792129Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6792318Z Traceback (most recent call last):
2025-12-04T09:39:16.6792533Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6792741Z     self.common(
2025-12-04T09:39:16.6792894Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6793107Z     return func(*args, **kwds)
2025-12-04T09:39:16.6793226Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6793428Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 725, in check_model_gpu
2025-12-04T09:39:16.6793636Z     check_model(
2025-12-04T09:39:16.6793817Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6794018Z     assert_equal_fn(
2025-12-04T09:39:16.6794226Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6794470Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6794617Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6794866Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6795185Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6795353Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6795443Z 
2025-12-04T09:39:16.6795496Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6795678Z Greatest absolute difference: 0.58544921875 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6795908Z Greatest relative difference: 0.57568359375 at index (0, 1) (up to 0.001 allowed)
2025-12-04T09:39:16.6796040Z 
2025-12-04T09:39:16.6796092Z The failure occurred for item [2]
2025-12-04T09:39:16.6796170Z 
2025-12-04T09:39:16.6796251Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6796578Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6796824Z 
2025-12-04T09:39:16.6796919Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6797130Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6797304Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6797722Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6798177Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6798356Z graph_break []
2025-12-04T09:39:16.6798490Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6798664Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6798863Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6799368Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6799803Z graph_break []
2025-12-04T09:39:16.6799937Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6800112Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6800310Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6800843Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6801282Z graph_break []
2025-12-04T09:39:16.6801418Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6801627Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6801824Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6802313Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6802738Z graph_break []
2025-12-04T09:39:16.6802872Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6803048Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6803243Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6803773Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6804205Z graph_break []
2025-12-04T09:39:16.6804339Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6804513Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6804712Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6805205Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6805638Z graph_break []
2025-12-04T09:39:16.6805771Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6805949Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6806145Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6806634Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6807060Z graph_break []
2025-12-04T09:39:16.6807193Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6807476Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6807673Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6808170Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6808599Z graph_break []
2025-12-04T09:39:16.6808731Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6808905Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6809102Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6809591Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6810019Z graph_break []
2025-12-04T09:39:16.6810151Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6810355Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6810587Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6811080Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6811510Z graph_break []
2025-12-04T09:39:16.6811642Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6811818Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6812015Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6812553Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6812982Z graph_break []
2025-12-04T09:39:16.6813114Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6813289Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6813487Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6813972Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6814401Z graph_break []
2025-12-04T09:39:16.6814534Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6814711Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6814908Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6815398Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6815823Z graph_break []
2025-12-04T09:39:16.6815957Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6816137Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6816332Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6816835Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6817263Z graph_break []
2025-12-04T09:39:16.6817396Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6817568Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6817764Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6818249Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6818678Z graph_break []
2025-12-04T09:39:16.6818813Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6819019Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6819216Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6819700Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6820129Z graph_break []
2025-12-04T09:39:16.6820261Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6820471Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6820668Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6821195Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6821623Z graph_break []
2025-12-04T09:39:16.6821757Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6821932Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6822128Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6822619Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6823044Z graph_break []
2025-12-04T09:39:16.6823180Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6823356Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6823552Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6824042Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6824472Z graph_break []
2025-12-04T09:39:16.6829901Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6830089Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6830289Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6830851Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6831281Z graph_break []
2025-12-04T09:39:16.6831445Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6831639Z Traceback (most recent call last):
2025-12-04T09:39:16.6831861Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6832072Z     self.common(
2025-12-04T09:39:16.6832228Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6832409Z     return func(*args, **kwds)
2025-12-04T09:39:16.6832527Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6832732Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.6832942Z     check_model(
2025-12-04T09:39:16.6833177Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6833375Z     assert_equal_fn(
2025-12-04T09:39:16.6833589Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6833836Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6833984Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6834235Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6834516Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6834683Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6834775Z 
2025-12-04T09:39:16.6834830Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6835021Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6835310Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.6835453Z 
2025-12-04T09:39:16.6835505Z The failure occurred for item [2]
2025-12-04T09:39:16.6835585Z 
2025-12-04T09:39:16.6835668Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6836005Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6836259Z 
2025-12-04T09:39:16.6836355Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6836567Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6836747Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6837179Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6837642Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6837823Z graph_break []
2025-12-04T09:39:16.6837958Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6838130Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6838318Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6838817Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6839250Z graph_break []
2025-12-04T09:39:16.6839389Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6839565Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6839762Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6840253Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6840721Z graph_break []
2025-12-04T09:39:16.6840852Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6841029Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6841225Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6841760Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6842181Z graph_break []
2025-12-04T09:39:16.6842315Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6842490Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6842682Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6843171Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6843603Z graph_break []
2025-12-04T09:39:16.6843767Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6843944Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6844141Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6844631Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6845062Z graph_break []
2025-12-04T09:39:16.6845198Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6845372Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6845566Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6846064Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6846493Z graph_break []
2025-12-04T09:39:16.6846628Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6846804Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6847002Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6847503Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6847934Z graph_break []
2025-12-04T09:39:16.6848068Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6848247Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6848446Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6848942Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6849383Z graph_break []
2025-12-04T09:39:16.6849514Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6849692Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6849890Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6850462Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6850892Z graph_break []
2025-12-04T09:39:16.6851024Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6851204Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6851403Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6851895Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6852326Z graph_break []
2025-12-04T09:39:16.6852500Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6852674Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6852870Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6853359Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6853784Z graph_break []
2025-12-04T09:39:16.6853917Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6854085Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6854277Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6854764Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6855193Z graph_break []
2025-12-04T09:39:16.6855323Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6855495Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6855688Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6856177Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6856603Z graph_break []
2025-12-04T09:39:16.6856735Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6856908Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6857100Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6857583Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6858009Z graph_break []
2025-12-04T09:39:16.6858137Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6858310Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6858506Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6859020Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6859445Z graph_break []
2025-12-04T09:39:16.6859576Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6859752Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6859945Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6860473Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6860902Z graph_break []
2025-12-04T09:39:16.6861064Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6861240Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6861434Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6861919Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6862348Z graph_break []
2025-12-04T09:39:16.6862479Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6862652Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6862846Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6863336Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6863762Z graph_break []
2025-12-04T09:39:16.6863892Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6864065Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6864259Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6864745Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6865173Z graph_break []
2025-12-04T09:39:16.6865306Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6865478Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6865667Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6866157Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6866587Z graph_break []
2025-12-04T09:39:16.6866719Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6866893Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6867088Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6867605Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6868027Z graph_break []
2025-12-04T09:39:16.6868186Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6868370Z Traceback (most recent call last):
2025-12-04T09:39:16.6868580Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6868783Z     self.common(
2025-12-04T09:39:16.6868931Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6869104Z     return func(*args, **kwds)
2025-12-04T09:39:16.6869218Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6869418Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.6869654Z     check_model(
2025-12-04T09:39:16.6869835Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6870034Z     assert_equal_fn(
2025-12-04T09:39:16.6870240Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6870523Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6870667Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6870915Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6871191Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6871355Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6871444Z 
2025-12-04T09:39:16.6871494Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6871679Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6871926Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.6872071Z 
2025-12-04T09:39:16.6872119Z The failure occurred for item [2]
2025-12-04T09:39:16.6872198Z 
2025-12-04T09:39:16.6872277Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6872602Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6872850Z 
2025-12-04T09:39:16.6872943Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6873144Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6873316Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6873731Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6874187Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6874364Z graph_break []
2025-12-04T09:39:16.6874495Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6874673Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6874867Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6875355Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6875779Z graph_break []
2025-12-04T09:39:16.6875911Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6876135Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6876328Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6876814Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6877240Z graph_break []
2025-12-04T09:39:16.6877369Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6877542Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6877735Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6878225Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6878689Z graph_break []
2025-12-04T09:39:16.6878818Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6878989Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6879180Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6879665Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6880090Z graph_break []
2025-12-04T09:39:16.6880223Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6880398Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6880634Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6881127Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6881553Z graph_break []
2025-12-04T09:39:16.6881684Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6881855Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6882049Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6882537Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6882962Z graph_break []
2025-12-04T09:39:16.6883092Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6883266Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6883459Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6883958Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6884387Z graph_break []
2025-12-04T09:39:16.6884517Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6884691Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6884923Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6885422Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6885847Z graph_break []
2025-12-04T09:39:16.6885978Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6886149Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6886343Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6886827Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6887287Z graph_break []
2025-12-04T09:39:16.6887416Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6887587Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6887780Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6888268Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6888692Z graph_break []
2025-12-04T09:39:16.6888822Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6888999Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6889192Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6889681Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6890102Z graph_break []
2025-12-04T09:39:16.6890233Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6890400Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6890623Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6891107Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6891535Z graph_break []
2025-12-04T09:39:16.6891665Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6891839Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6892034Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6892519Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6892941Z graph_break []
2025-12-04T09:39:16.6893070Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6893246Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6893470Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6893952Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6894375Z graph_break []
2025-12-04T09:39:16.6894506Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6894679Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6894872Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6895362Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6895816Z graph_break []
2025-12-04T09:39:16.6895945Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6896119Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6896315Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6896809Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6897242Z graph_break []
2025-12-04T09:39:16.6897372Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6897550Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6897747Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6898234Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6898663Z graph_break []
2025-12-04T09:39:16.6898794Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6898965Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6899155Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6899653Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6900086Z graph_break []
2025-12-04T09:39:16.6900216Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6900388Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6900623Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6901110Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6901533Z graph_break []
2025-12-04T09:39:16.6901664Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6901839Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6902066Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6902551Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6902977Z graph_break []
2025-12-04T09:39:16.6903107Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6903278Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6903473Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6903959Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6904434Z graph_break []
2025-12-04T09:39:16.6904564Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6904734Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6904926Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6905410Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6905828Z graph_break []
2025-12-04T09:39:16.6905983Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6906172Z Traceback (most recent call last):
2025-12-04T09:39:16.6906385Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6906589Z     self.common(
2025-12-04T09:39:16.6906738Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6906911Z     return func(*args, **kwds)
2025-12-04T09:39:16.6907026Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6907227Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.6907434Z     check_model(
2025-12-04T09:39:16.6907612Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6907810Z     assert_equal_fn(
2025-12-04T09:39:16.6908014Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6908251Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6908528Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6908775Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6909045Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6909205Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6909295Z 
2025-12-04T09:39:16.6909341Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6909523Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6909766Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.6909907Z 
2025-12-04T09:39:16.6909955Z The failure occurred for item [2]
2025-12-04T09:39:16.6910032Z 
2025-12-04T09:39:16.6910108Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6910516Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6910764Z 
2025-12-04T09:39:16.6910854Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6911055Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6911224Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6911637Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6911737Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6911773Z graph_break []
2025-12-04T09:39:16.6911848Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6911938Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6912040Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6912394Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6912434Z graph_break []
2025-12-04T09:39:16.6912507Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6912565Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6912662Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6913018Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6913057Z graph_break []
2025-12-04T09:39:16.6913133Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6913188Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6913287Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6913637Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6913673Z graph_break []
2025-12-04T09:39:16.6913747Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6913804Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6913903Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6914250Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6914288Z graph_break []
2025-12-04T09:39:16.6914360Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6914417Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6914513Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6914890Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6914929Z graph_break []
2025-12-04T09:39:16.6915003Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6915057Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6915155Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6915511Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6915548Z graph_break []
2025-12-04T09:39:16.6915622Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6915698Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6915798Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6916156Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6916194Z graph_break []
2025-12-04T09:39:16.6916267Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6916324Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6916419Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6916775Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6916813Z graph_break []
2025-12-04T09:39:16.6916887Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6916942Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6917044Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6917397Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6917432Z graph_break []
2025-12-04T09:39:16.6917507Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6917564Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6917664Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6918017Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6918056Z graph_break []
2025-12-04T09:39:16.6918129Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6918186Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6918282Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6918658Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6918696Z graph_break []
2025-12-04T09:39:16.6918771Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6918826Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6918932Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6919285Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6919325Z graph_break []
2025-12-04T09:39:16.6919399Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6919482Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6919582Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6919939Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6919977Z graph_break []
2025-12-04T09:39:16.6920054Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6920109Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6920212Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6920603Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6920646Z graph_break []
2025-12-04T09:39:16.6920722Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6920779Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6920880Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6921232Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6921274Z graph_break []
2025-12-04T09:39:16.6921349Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6921410Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6921510Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6921864Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6921902Z graph_break []
2025-12-04T09:39:16.6921980Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6922036Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6922138Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6922521Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6922563Z graph_break []
2025-12-04T09:39:16.6922640Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6922697Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6922796Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6923148Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6923189Z graph_break []
2025-12-04T09:39:16.6923263Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6923360Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6923459Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6923820Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6923857Z graph_break []
2025-12-04T09:39:16.6923934Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6923990Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6924091Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6924444Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6924486Z graph_break []
2025-12-04T09:39:16.6924562Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6924618Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6924718Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6925071Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6925112Z graph_break []
2025-12-04T09:39:16.6925185Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6925246Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6925347Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6925700Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6925738Z graph_break []
2025-12-04T09:39:16.6925818Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6925874Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6925976Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6926349Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6926397Z graph_break []
2025-12-04T09:39:16.6926474Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6926532Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6926630Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6926986Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6927028Z graph_break []
2025-12-04T09:39:16.6927099Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6927184Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6927283Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6927637Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6927674Z graph_break []
2025-12-04T09:39:16.6927778Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6927826Z Traceback (most recent call last):
2025-12-04T09:39:16.6927962Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6928000Z     self.common(
2025-12-04T09:39:16.6928092Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6928138Z     return func(*args, **kwds)
2025-12-04T09:39:16.6928182Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6928314Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.6928355Z     check_model(
2025-12-04T09:39:16.6928476Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6928520Z     assert_equal_fn(
2025-12-04T09:39:16.6928666Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6928728Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6928778Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6928944Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6929021Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6929076Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6929079Z 
2025-12-04T09:39:16.6929130Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6929234Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6929344Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.6929346Z 
2025-12-04T09:39:16.6929392Z The failure occurred for item [2]
2025-12-04T09:39:16.6929393Z 
2025-12-04T09:39:16.6929472Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6929686Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6929688Z 
2025-12-04T09:39:16.6929780Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6929855Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6929915Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6930265Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6930365Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6930495Z graph_break []
2025-12-04T09:39:16.6930569Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6930631Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6930729Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6931090Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6931163Z graph_break []
2025-12-04T09:39:16.6931240Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6931298Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6931399Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6931752Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6931793Z graph_break []
2025-12-04T09:39:16.6931867Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6931931Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6932029Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6932394Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6932436Z graph_break []
2025-12-04T09:39:16.6932510Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6932572Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6932670Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6933026Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6933068Z graph_break []
2025-12-04T09:39:16.6933144Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6933201Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6933301Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6933651Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6933691Z graph_break []
2025-12-04T09:39:16.6933764Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6933823Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6933953Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6934305Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6934347Z graph_break []
2025-12-04T09:39:16.6934419Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6934481Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6934577Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6934933Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6934998Z graph_break []
2025-12-04T09:39:16.6935077Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6935134Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6935235Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6935589Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6935629Z graph_break []
2025-12-04T09:39:16.6935702Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6935763Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6935862Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6936216Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6936259Z graph_break []
2025-12-04T09:39:16.6936335Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6936394Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6936493Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6936854Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6936896Z graph_break []
2025-12-04T09:39:16.6936972Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6937026Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6937129Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6937477Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6937518Z graph_break []
2025-12-04T09:39:16.6937589Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6937649Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6937771Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6938123Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6938165Z graph_break []
2025-12-04T09:39:16.6938238Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6938300Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6938398Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6938758Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6938820Z graph_break []
2025-12-04T09:39:16.6938898Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6938953Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6939053Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6939401Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6939443Z graph_break []
2025-12-04T09:39:16.6939515Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6939573Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6939676Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6940026Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6940067Z graph_break []
2025-12-04T09:39:16.6940140Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6940200Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6940298Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6940694Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6940734Z graph_break []
2025-12-04T09:39:16.6940810Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6940867Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6940968Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6941323Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6941365Z graph_break []
2025-12-04T09:39:16.6941437Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6941499Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6941627Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6941978Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6942020Z graph_break []
2025-12-04T09:39:16.6942094Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6942155Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6942252Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6942606Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6942673Z graph_break []
2025-12-04T09:39:16.6942749Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6942805Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6942907Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6943260Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6943303Z graph_break []
2025-12-04T09:39:16.6943379Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6943436Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6943539Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6943889Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6943930Z graph_break []
2025-12-04T09:39:16.6944006Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6944066Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6944164Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6944515Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6944556Z graph_break []
2025-12-04T09:39:16.6944634Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6944691Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6944790Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6945143Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6945184Z graph_break []
2025-12-04T09:39:16.6945260Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6945316Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6945441Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6945792Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6945831Z graph_break []
2025-12-04T09:39:16.6945905Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6945964Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6946062Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6946416Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6946480Z graph_break []
2025-12-04T09:39:16.6946557Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6946614Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6946715Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6947065Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6947106Z graph_break []
2025-12-04T09:39:16.6947199Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6947255Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6947361Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6947711Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6947751Z graph_break []
2025-12-04T09:39:16.6947825Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6947885Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6947985Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6948345Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6948385Z graph_break []
2025-12-04T09:39:16.6948460Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6948515Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6948617Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6948973Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6949015Z graph_break []
2025-12-04T09:39:16.6949121Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6949168Z Traceback (most recent call last):
2025-12-04T09:39:16.6949306Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6949366Z     self.common(
2025-12-04T09:39:16.6949463Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6949507Z     return func(*args, **kwds)
2025-12-04T09:39:16.6949550Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6949680Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.6949723Z     check_model(
2025-12-04T09:39:16.6949842Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6949885Z     assert_equal_fn(
2025-12-04T09:39:16.6950027Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6950093Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6950165Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6950332Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6950446Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6950505Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6950507Z 
2025-12-04T09:39:16.6950552Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6950660Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6950763Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.6950766Z 
2025-12-04T09:39:16.6950815Z The failure occurred for item [2]
2025-12-04T09:39:16.6950817Z 
2025-12-04T09:39:16.6950893Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6951107Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6951111Z 
2025-12-04T09:39:16.6951203Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6951276Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6951337Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6951656Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6951759Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6951797Z graph_break []
2025-12-04T09:39:16.6951874Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6951933Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6952037Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6952390Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6952432Z graph_break []
2025-12-04T09:39:16.6952504Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6952565Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6952662Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6953021Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6953100Z graph_break []
2025-12-04T09:39:16.6953175Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6953236Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6953335Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6953690Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6953729Z graph_break []
2025-12-04T09:39:16.6953808Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6953864Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6954000Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6954347Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6954388Z graph_break []
2025-12-04T09:39:16.6954462Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6954523Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6954621Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6954971Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6955016Z graph_break []
2025-12-04T09:39:16.6955090Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6955151Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6955250Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6955601Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6955638Z graph_break []
2025-12-04T09:39:16.6955715Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6955771Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6955875Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6956232Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6956273Z graph_break []
2025-12-04T09:39:16.6956346Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6956406Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6956505Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6956862Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6956963Z graph_break []
2025-12-04T09:39:16.6957036Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6957097Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6957195Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6957550Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6957587Z graph_break []
2025-12-04T09:39:16.6957664Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6957720Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6957847Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6958200Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6958241Z graph_break []
2025-12-04T09:39:16.6958315Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6958374Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6958471Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6958822Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6958865Z graph_break []
2025-12-04T09:39:16.6958939Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6958999Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6959098Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6959451Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6959489Z graph_break []
2025-12-04T09:39:16.6959567Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6959626Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6959726Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6960077Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6960116Z graph_break []
2025-12-04T09:39:16.6960189Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6960245Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6960342Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6960727Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6960796Z graph_break []
2025-12-04T09:39:16.6960869Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6960925Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6961021Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6961368Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6961404Z graph_break []
2025-12-04T09:39:16.6961478Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6961533Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6961662Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6962015Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6962053Z graph_break []
2025-12-04T09:39:16.6962125Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6962183Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6962280Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6962644Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6962687Z graph_break []
2025-12-04T09:39:16.6962760Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6962817Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6962913Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6963267Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6963304Z graph_break []
2025-12-04T09:39:16.6963380Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6963436Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6963536Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6963887Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6963926Z graph_break []
2025-12-04T09:39:16.6963998Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6964056Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6964155Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6964508Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6964549Z graph_break []
2025-12-04T09:39:16.6964646Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6964703Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6964800Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6965159Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6965195Z graph_break []
2025-12-04T09:39:16.6965275Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6965330Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6965466Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6965818Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6965858Z graph_break []
2025-12-04T09:39:16.6965933Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6965997Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6966097Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6966447Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6966486Z graph_break []
2025-12-04T09:39:16.6966560Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6966621Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6966720Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6967077Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6967113Z graph_break []
2025-12-04T09:39:16.6967187Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6967243Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6967344Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6967690Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6967730Z graph_break []
2025-12-04T09:39:16.6967801Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6967860Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6967959Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6968312Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6968353Z graph_break []
2025-12-04T09:39:16.6968448Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6968507Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6968603Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6968962Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6968998Z graph_break []
2025-12-04T09:39:16.6969072Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6969128Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6969252Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6969604Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6969644Z graph_break []
2025-12-04T09:39:16.6969716Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6969774Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6969874Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6970220Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6970260Z graph_break []
2025-12-04T09:39:16.6970334Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6970391Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6970533Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6970893Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6970929Z graph_break []
2025-12-04T09:39:16.6971032Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6971079Z Traceback (most recent call last):
2025-12-04T09:39:16.6971212Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6971251Z     self.common(
2025-12-04T09:39:16.6971345Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6971389Z     return func(*args, **kwds)
2025-12-04T09:39:16.6971431Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6971560Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 725, in check_model_gpu
2025-12-04T09:39:16.6971599Z     check_model(
2025-12-04T09:39:16.6971718Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6971758Z     assert_equal_fn(
2025-12-04T09:39:16.6971899Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6971962Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6972007Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6972218Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6972295Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6972348Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6972350Z 
2025-12-04T09:39:16.6972398Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6972495Z Greatest absolute difference: 0.58544921875 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6972591Z Greatest relative difference: 0.57568359375 at index (0, 1) (up to 0.001 allowed)
2025-12-04T09:39:16.6972593Z 
2025-12-04T09:39:16.6972638Z The failure occurred for item [2]
2025-12-04T09:39:16.6972640Z 
2025-12-04T09:39:16.6972722Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6972936Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6972968Z 
2025-12-04T09:39:16.6973069Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6973143Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6973202Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6973525Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6973628Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6973667Z graph_break []
2025-12-04T09:39:16.6973741Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6973799Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6973898Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6974256Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6974293Z graph_break []
2025-12-04T09:39:16.6974368Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6974427Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6974529Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6974887Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6974931Z graph_break []
2025-12-04T09:39:16.6975005Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6975066Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6975163Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6975521Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6975561Z graph_break []
2025-12-04T09:39:16.6975634Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6975698Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6975799Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6976179Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6976216Z graph_break []
2025-12-04T09:39:16.6976292Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6976350Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6976452Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6976802Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6976873Z graph_break []
2025-12-04T09:39:16.6976947Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6977008Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6977106Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6977459Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6977503Z graph_break []
2025-12-04T09:39:16.6977578Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6977638Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6977736Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6978096Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6978134Z graph_break []
2025-12-04T09:39:16.6978211Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6978266Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6978367Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6978717Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6978758Z graph_break []
2025-12-04T09:39:16.6978832Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6978891Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6978988Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6979340Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6979382Z graph_break []
2025-12-04T09:39:16.6979454Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6979513Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6979610Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6979985Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6980022Z graph_break []
2025-12-04T09:39:16.6980095Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6980151Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6980250Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6980643Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6980717Z graph_break []
2025-12-04T09:39:16.6980791Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6980847Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6980944Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6981293Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6981332Z graph_break []
2025-12-04T09:39:16.6981404Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6981460Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6981556Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6981919Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6981955Z graph_break []
2025-12-04T09:39:16.6982029Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6982083Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6982180Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6982530Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6982570Z graph_break []
2025-12-04T09:39:16.6982643Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6982699Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6982795Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6983144Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6983182Z graph_break []
2025-12-04T09:39:16.6983254Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6983313Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6983410Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6983797Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6983834Z graph_break []
2025-12-04T09:39:16.6983909Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6983964Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6984063Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6984414Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6984477Z graph_break []
2025-12-04T09:39:16.6984551Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6984610Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6984706Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6985060Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6985098Z graph_break []
2025-12-04T09:39:16.6985169Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6985226Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6985323Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6985679Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6985716Z graph_break []
2025-12-04T09:39:16.6985789Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6985843Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6985942Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6986295Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6986336Z graph_break []
2025-12-04T09:39:16.6986409Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6986464Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6986560Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6986911Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6986949Z graph_break []
2025-12-04T09:39:16.6987022Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6987077Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6987173Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6987551Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6987587Z graph_break []
2025-12-04T09:39:16.6987660Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6987716Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6987813Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6988161Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6988228Z graph_break []
2025-12-04T09:39:16.6988302Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6988360Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6988456Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6988807Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6988846Z graph_break []
2025-12-04T09:39:16.6988917Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6988973Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6989069Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6989423Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6989460Z graph_break []
2025-12-04T09:39:16.6989533Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6989588Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6989686Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6990034Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6990075Z graph_break []
2025-12-04T09:39:16.6990148Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6990205Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6990303Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6990703Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6990742Z graph_break []
2025-12-04T09:39:16.6990814Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6990870Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6990967Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6991362Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6991398Z graph_break []
2025-12-04T09:39:16.6991472Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6991526Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6991624Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6991971Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6992046Z graph_break []
2025-12-04T09:39:16.6992119Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6992174Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6992271Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6992629Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6992666Z graph_break []
2025-12-04T09:39:16.6992737Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6992796Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6992893Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6993248Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6993284Z graph_break []
2025-12-04T09:39:16.6993384Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.6993432Z Traceback (most recent call last):
2025-12-04T09:39:16.6993563Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.6993600Z     self.common(
2025-12-04T09:39:16.6993690Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.6993732Z     return func(*args, **kwds)
2025-12-04T09:39:16.6993772Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6993901Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.6993944Z     check_model(
2025-12-04T09:39:16.6994065Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.6994105Z     assert_equal_fn(
2025-12-04T09:39:16.6994247Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.6994309Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.6994353Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.6994518Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.6994590Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.6994645Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.6994647Z 
2025-12-04T09:39:16.6994693Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.6994797Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.6994932Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.6994934Z 
2025-12-04T09:39:16.6994980Z The failure occurred for item [2]
2025-12-04T09:39:16.6994982Z 
2025-12-04T09:39:16.6995059Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.6995269Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.6995272Z 
2025-12-04T09:39:16.6995361Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.6995434Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6995493Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6995811Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6995937Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6995973Z graph_break []
2025-12-04T09:39:16.6996048Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6996105Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6996204Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6996555Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6996593Z graph_break []
2025-12-04T09:39:16.6996670Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6996727Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6996826Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6997177Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6997215Z graph_break []
2025-12-04T09:39:16.6997287Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6997344Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6997439Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6997794Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6997831Z graph_break []
2025-12-04T09:39:16.6997905Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6997963Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6998059Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6998410Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6998448Z graph_break []
2025-12-04T09:39:16.6998542Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6998599Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6998697Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.6999050Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.6999088Z graph_break []
2025-12-04T09:39:16.6999160Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6999216Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.6999312Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.6999684Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.6999722Z graph_break []
2025-12-04T09:39:16.6999794Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.6999853Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.6999949Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7000304Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7000342Z graph_break []
2025-12-04T09:39:16.7000459Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7000515Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7003961Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7004324Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7004368Z graph_break []
2025-12-04T09:39:16.7004450Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7004508Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7004614Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7004976Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7005020Z graph_break []
2025-12-04T09:39:16.7005093Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7005153Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7005251Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7005609Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7005649Z graph_break []
2025-12-04T09:39:16.7005774Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7005831Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7005935Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7006286Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7006329Z graph_break []
2025-12-04T09:39:16.7006406Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7006462Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7006564Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7006948Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7006992Z graph_break []
2025-12-04T09:39:16.7007067Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7007129Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7007228Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7007585Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7007625Z graph_break []
2025-12-04T09:39:16.7007706Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7007764Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7007865Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7008215Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7008253Z graph_break []
2025-12-04T09:39:16.7008332Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7008388Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7008488Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7008842Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7008884Z graph_break []
2025-12-04T09:39:16.7008958Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7009021Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7009121Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7009478Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7009518Z graph_break []
2025-12-04T09:39:16.7009596Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7009677Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7009781Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7010134Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7010173Z graph_break []
2025-12-04T09:39:16.7010251Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7010309Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7010451Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7010830Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7010873Z graph_break []
2025-12-04T09:39:16.7010948Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7011011Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7011107Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7011462Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7011502Z graph_break []
2025-12-04T09:39:16.7011585Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7011645Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7011748Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7012104Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7012145Z graph_break []
2025-12-04T09:39:16.7012224Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7012283Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7012383Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7012735Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7012779Z graph_break []
2025-12-04T09:39:16.7012855Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7012913Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7013009Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7013451Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7013489Z graph_break []
2025-12-04T09:39:16.7013567Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7013666Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7013769Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7014124Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7014162Z graph_break []
2025-12-04T09:39:16.7014241Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7014299Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7014402Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7014783Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7014827Z graph_break []
2025-12-04T09:39:16.7014903Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7014963Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7015062Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7015408Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7015449Z graph_break []
2025-12-04T09:39:16.7015527Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7015586Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7015689Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7016043Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7016082Z graph_break []
2025-12-04T09:39:16.7016161Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7016218Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7016319Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7016673Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7016717Z graph_break []
2025-12-04T09:39:16.7016792Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7016854Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7016952Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7017302Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7017341Z graph_break []
2025-12-04T09:39:16.7017421Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7017500Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7017603Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7017955Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7017993Z graph_break []
2025-12-04T09:39:16.7018071Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7018128Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7018228Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7018604Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7018647Z graph_break []
2025-12-04T09:39:16.7018724Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7018788Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7018886Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7019243Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7019282Z graph_break []
2025-12-04T09:39:16.7019364Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7019424Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7019529Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7019883Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7019923Z graph_break []
2025-12-04T09:39:16.7020002Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7020058Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7020161Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7020547Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7020590Z graph_break []
2025-12-04T09:39:16.7020666Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7020729Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7020826Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7021183Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7021222Z graph_break []
2025-12-04T09:39:16.7021304Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7021391Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7021493Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7021849Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7021886Z graph_break []
2025-12-04T09:39:16.7021992Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.7022042Z Traceback (most recent call last):
2025-12-04T09:39:16.7022182Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.7022223Z     self.common(
2025-12-04T09:39:16.7022356Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.7022405Z     return func(*args, **kwds)
2025-12-04T09:39:16.7022451Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.7022586Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 725, in check_model_gpu
2025-12-04T09:39:16.7022629Z     check_model(
2025-12-04T09:39:16.7022751Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.7022796Z     assert_equal_fn(
2025-12-04T09:39:16.7022944Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.7023012Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.7023060Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.7023233Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.7023310Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.7023373Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7023376Z 
2025-12-04T09:39:16.7023424Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7023525Z Greatest absolute difference: 0.58544921875 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7023622Z Greatest relative difference: 0.57568359375 at index (0, 1) (up to 0.001 allowed)
2025-12-04T09:39:16.7023624Z 
2025-12-04T09:39:16.7023677Z The failure occurred for item [2]
2025-12-04T09:39:16.7023679Z 
2025-12-04T09:39:16.7023762Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7023978Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7023980Z 
2025-12-04T09:39:16.7024075Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7024152Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7024216Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7024536Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7024641Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7024680Z graph_break []
2025-12-04T09:39:16.7024759Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7024817Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7024921Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7025296Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7025340Z graph_break []
2025-12-04T09:39:16.7025414Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7025475Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7025576Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7025937Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7025974Z graph_break []
2025-12-04T09:39:16.7026085Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7026144Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7026243Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7026595Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7026635Z graph_break []
2025-12-04T09:39:16.7026707Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7026767Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7026863Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7027219Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7027257Z graph_break []
2025-12-04T09:39:16.7027333Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7027391Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7027490Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7027847Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7027885Z graph_break []
2025-12-04T09:39:16.7027963Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7028022Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7028120Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7028469Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7028512Z graph_break []
2025-12-04T09:39:16.7028585Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7028645Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7028742Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7029116Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7029155Z graph_break []
2025-12-04T09:39:16.7029231Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7029289Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7029389Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7029742Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7029780Z graph_break []
2025-12-04T09:39:16.7029882Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7029941Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7030041Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7030392Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7030466Z graph_break []
2025-12-04T09:39:16.7030540Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7030602Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7030699Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7031057Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7031096Z graph_break []
2025-12-04T09:39:16.7031175Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7031233Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7031333Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7031681Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7031720Z graph_break []
2025-12-04T09:39:16.7031799Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7031857Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7031958Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7032305Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7032348Z graph_break []
2025-12-04T09:39:16.7032421Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7032482Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7032579Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7032957Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7032997Z graph_break []
2025-12-04T09:39:16.7033075Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7033133Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7033229Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7033578Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7033616Z graph_break []
2025-12-04T09:39:16.7033729Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7033786Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7033885Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7034232Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7034273Z graph_break []
2025-12-04T09:39:16.7034347Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7034406Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7034505Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7034863Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7034901Z graph_break []
2025-12-04T09:39:16.7034977Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7035037Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7035134Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7035484Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7035521Z graph_break []
2025-12-04T09:39:16.7035597Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7035657Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7035756Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7036108Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7036147Z graph_break []
2025-12-04T09:39:16.7036220Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7036281Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7036377Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7036758Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7036800Z graph_break []
2025-12-04T09:39:16.7036874Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7036934Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7037032Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7037395Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7037432Z graph_break []
2025-12-04T09:39:16.7037510Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7037589Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7037692Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7038041Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7038083Z graph_break []
2025-12-04T09:39:16.7038156Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7038216Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7038313Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7038663Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7038707Z graph_break []
2025-12-04T09:39:16.7038781Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7038844Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7038941Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7039295Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7039332Z graph_break []
2025-12-04T09:39:16.7039409Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7039469Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7039571Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7039920Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7039963Z graph_break []
2025-12-04T09:39:16.7040035Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7040092Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7040189Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7040595Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7040637Z graph_break []
2025-12-04T09:39:16.7040712Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7040772Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7040869Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7041218Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7041257Z graph_break []
2025-12-04T09:39:16.7041333Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7041426Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7041528Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7041876Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7041917Z graph_break []
2025-12-04T09:39:16.7041992Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7042053Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7042149Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7042500Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7042543Z graph_break []
2025-12-04T09:39:16.7042616Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7042674Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7042771Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7043206Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7043245Z graph_break []
2025-12-04T09:39:16.7043321Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7043376Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7043478Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7043823Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7043865Z graph_break []
2025-12-04T09:39:16.7043938Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7043997Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7044092Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7044471Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7044515Z graph_break []
2025-12-04T09:39:16.7044587Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7044646Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7044744Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7045097Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7045134Z graph_break []
2025-12-04T09:39:16.7045212Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7045293Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7045394Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7045742Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7045784Z graph_break []
2025-12-04T09:39:16.7045857Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7045917Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7046013Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7046366Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7046410Z graph_break []
2025-12-04T09:39:16.7046483Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7046542Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7046638Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7046986Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7047024Z graph_break []
2025-12-04T09:39:16.7047102Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7047159Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7047262Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7047612Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7047652Z graph_break []
2025-12-04T09:39:16.7047727Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7047786Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7047882Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7048256Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7048300Z graph_break []
2025-12-04T09:39:16.7048401Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.7048453Z Traceback (most recent call last):
2025-12-04T09:39:16.7048585Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.7048630Z     self.common(
2025-12-04T09:39:16.7048721Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.7048769Z     return func(*args, **kwds)
2025-12-04T09:39:16.7048810Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.7048944Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.7048983Z     check_model(
2025-12-04T09:39:16.7049132Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.7049174Z     assert_equal_fn(
2025-12-04T09:39:16.7049320Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.7049381Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.7049430Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.7049595Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.7049671Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.7049726Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7049728Z 
2025-12-04T09:39:16.7049778Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7049881Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7049991Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7049993Z 
2025-12-04T09:39:16.7050043Z The failure occurred for item [2]
2025-12-04T09:39:16.7050045Z 
2025-12-04T09:39:16.7050127Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7050344Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7050346Z 
2025-12-04T09:39:16.7050466Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7050546Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7050605Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7050929Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7051032Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7051074Z graph_break []
2025-12-04T09:39:16.7051148Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7051210Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7051307Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7051661Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7051699Z graph_break []
2025-12-04T09:39:16.7051778Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7051836Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7051965Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7052318Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7052356Z graph_break []
2025-12-04T09:39:16.7052434Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7052490Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7052592Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7052940Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7053012Z graph_break []
2025-12-04T09:39:16.7053086Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7053146Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7053242Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7053593Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7053630Z graph_break []
2025-12-04T09:39:16.7053709Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7053770Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7053870Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7054224Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7054261Z graph_break []
2025-12-04T09:39:16.7054336Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7054392Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7054491Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7054843Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7054886Z graph_break []
2025-12-04T09:39:16.7054958Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7055017Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7055112Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7055460Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7055497Z graph_break []
2025-12-04T09:39:16.7055573Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7055631Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7055760Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7056115Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7056152Z graph_break []
2025-12-04T09:39:16.7056228Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7056284Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7056384Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7056732Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7056798Z graph_break []
2025-12-04T09:39:16.7056870Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7056929Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7057024Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7057378Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7057416Z graph_break []
2025-12-04T09:39:16.7057492Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7057550Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7057653Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7058003Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7058042Z graph_break []
2025-12-04T09:39:16.7058118Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7058173Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7058273Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7058632Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7058674Z graph_break []
2025-12-04T09:39:16.7058747Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7058808Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7058905Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7059258Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7059298Z graph_break []
2025-12-04T09:39:16.7059374Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7059433Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7059553Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7059905Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7059943Z graph_break []
2025-12-04T09:39:16.7060020Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7060076Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7060177Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7060561Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7060633Z graph_break []
2025-12-04T09:39:16.7060706Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7060766Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7060862Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7061216Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7061255Z graph_break []
2025-12-04T09:39:16.7061330Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7061394Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7061494Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7061843Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7061880Z graph_break []
2025-12-04T09:39:16.7061959Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7062015Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7062116Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7062463Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7062506Z graph_break []
2025-12-04T09:39:16.7062578Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7062637Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7062735Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7063086Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7063129Z graph_break []
2025-12-04T09:39:16.7063202Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7063265Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7063388Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7063741Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7063781Z graph_break []
2025-12-04T09:39:16.7063861Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7063917Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7064017Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7064364Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7064432Z graph_break []
2025-12-04T09:39:16.7064506Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7064565Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7064661Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7065009Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7065050Z graph_break []
2025-12-04T09:39:16.7065124Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7065186Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7065284Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7065636Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7065674Z graph_break []
2025-12-04T09:39:16.7065750Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7065806Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7065907Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7066258Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7066300Z graph_break []
2025-12-04T09:39:16.7066371Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7066432Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7066530Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7066878Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7066916Z graph_break []
2025-12-04T09:39:16.7066992Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7067053Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7067174Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7067524Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7067562Z graph_break []
2025-12-04T09:39:16.7067639Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7067697Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7067799Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7068149Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7068217Z graph_break []
2025-12-04T09:39:16.7068292Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7068352Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7068449Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7068799Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7068838Z graph_break []
2025-12-04T09:39:16.7068916Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7068972Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7069073Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7069421Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7069459Z graph_break []
2025-12-04T09:39:16.7069535Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7069590Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7069689Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7070041Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7070085Z graph_break []
2025-12-04T09:39:16.7070159Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7070218Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7070315Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7070706Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7070744Z graph_break []
2025-12-04T09:39:16.7070822Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7070880Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7071017Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7071371Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7071408Z graph_break []
2025-12-04T09:39:16.7071487Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7071543Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7071642Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7071992Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7072069Z graph_break []
2025-12-04T09:39:16.7072142Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7072201Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7072297Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7072650Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7072687Z graph_break []
2025-12-04T09:39:16.7072762Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7072818Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7072919Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7073270Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7073309Z graph_break []
2025-12-04T09:39:16.7073386Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7073443Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7073544Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7073894Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7073938Z graph_break []
2025-12-04T09:39:16.7074011Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7074072Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7074168Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7074522Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7074560Z graph_break []
2025-12-04T09:39:16.7074636Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7074698Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7074816Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7075166Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7075202Z graph_break []
2025-12-04T09:39:16.7075278Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7075333Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7075430Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7075778Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7075840Z graph_break []
2025-12-04T09:39:16.7075912Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7075969Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7076064Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7076413Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7076451Z graph_break []
2025-12-04T09:39:16.7076524Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7076581Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7076678Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7077026Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7077063Z graph_break []
2025-12-04T09:39:16.7077167Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.7077213Z Traceback (most recent call last):
2025-12-04T09:39:16.7077344Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.7077382Z     self.common(
2025-12-04T09:39:16.7077474Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.7077518Z     return func(*args, **kwds)
2025-12-04T09:39:16.7077558Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.7077688Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.7077727Z     check_model(
2025-12-04T09:39:16.7077845Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.7077885Z     assert_equal_fn(
2025-12-04T09:39:16.7078027Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.7078089Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.7078133Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.7078297Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.7078369Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.7078425Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7078427Z 
2025-12-04T09:39:16.7078471Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7078598Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7078700Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7078705Z 
2025-12-04T09:39:16.7078750Z The failure occurred for item [2]
2025-12-04T09:39:16.7078752Z 
2025-12-04T09:39:16.7078828Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7079037Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7079039Z 
2025-12-04T09:39:16.7079129Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7079202Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7079284Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7079603Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7079705Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7079741Z graph_break []
2025-12-04T09:39:16.7079816Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7079873Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7079973Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7080325Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7080365Z graph_break []
2025-12-04T09:39:16.7080479Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7080535Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7080633Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7080982Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7081021Z graph_break []
2025-12-04T09:39:16.7081093Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7081153Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7081251Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7081600Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7081636Z graph_break []
2025-12-04T09:39:16.7081712Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7081768Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7081866Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7082244Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7082286Z graph_break []
2025-12-04T09:39:16.7082360Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7082416Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7082513Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7082859Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7082898Z graph_break []
2025-12-04T09:39:16.7082969Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7083063Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7083162Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7083507Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7083543Z graph_break []
2025-12-04T09:39:16.7083619Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7083676Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7083774Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7084121Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7084165Z graph_break []
2025-12-04T09:39:16.7084239Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7084294Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7084392Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7084738Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7084776Z graph_break []
2025-12-04T09:39:16.7084848Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7084908Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7085006Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7085355Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7085394Z graph_break []
2025-12-04T09:39:16.7085466Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7085523Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7085618Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7085973Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7086038Z graph_break []
2025-12-04T09:39:16.7086111Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7086170Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7086267Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7086624Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7086661Z graph_break []
2025-12-04T09:39:16.7086735Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7086789Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7086913Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7087258Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7087296Z graph_break []
2025-12-04T09:39:16.7087368Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7087426Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7087525Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7087871Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7087914Z graph_break []
2025-12-04T09:39:16.7087987Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7088044Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7088139Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7088483Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7088519Z graph_break []
2025-12-04T09:39:16.7088593Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7088648Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7088748Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7089094Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7089134Z graph_break []
2025-12-04T09:39:16.7089205Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7089264Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7089364Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7089716Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7089781Z graph_break []
2025-12-04T09:39:16.7089853Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7089910Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7090005Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7090356Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7090392Z graph_break []
2025-12-04T09:39:16.7090495Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7090550Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7090681Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7091032Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7091070Z graph_break []
2025-12-04T09:39:16.7091142Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7091201Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7091298Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7091646Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7091686Z graph_break []
2025-12-04T09:39:16.7091757Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7091814Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7091910Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7092260Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7092297Z graph_break []
2025-12-04T09:39:16.7092371Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7092425Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7092525Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7092875Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7092916Z graph_break []
2025-12-04T09:39:16.7092991Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7093045Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7093142Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7093492Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7093557Z graph_break []
2025-12-04T09:39:16.7093630Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7093688Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7093783Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7094135Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7094171Z graph_break []
2025-12-04T09:39:16.7094247Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7094302Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7094424Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7094774Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7094813Z graph_break []
2025-12-04T09:39:16.7094887Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7094941Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7095038Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7095383Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7095425Z graph_break []
2025-12-04T09:39:16.7095496Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7095553Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7095648Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7095998Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7096033Z graph_break []
2025-12-04T09:39:16.7096107Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7096161Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7096260Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7096607Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7096645Z graph_break []
2025-12-04T09:39:16.7096718Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7096773Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7096871Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7097217Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7097278Z graph_break []
2025-12-04T09:39:16.7097351Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7097408Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7097503Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7097852Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7097888Z graph_break []
2025-12-04T09:39:16.7097961Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7098014Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7098145Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7098493Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7098531Z graph_break []
2025-12-04T09:39:16.7098607Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7098663Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7098761Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7099120Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7099160Z graph_break []
2025-12-04T09:39:16.7099233Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7099289Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7099385Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7099735Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7099771Z graph_break []
2025-12-04T09:39:16.7099845Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7099900Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7100000Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7100349Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7100389Z graph_break []
2025-12-04T09:39:16.7100494Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7100551Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7100652Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7100996Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7101039Z graph_break []
2025-12-04T09:39:16.7101142Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7101199Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7101296Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7101641Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7101677Z graph_break []
2025-12-04T09:39:16.7101753Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7101808Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7101942Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7102296Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7102334Z graph_break []
2025-12-04T09:39:16.7102408Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7102464Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7102564Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7102916Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7102957Z graph_break []
2025-12-04T09:39:16.7103034Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7103095Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7103190Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7103541Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7103577Z graph_break []
2025-12-04T09:39:16.7103653Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7103709Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7103811Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7104162Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7104200Z graph_break []
2025-12-04T09:39:16.7104279Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7104335Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7104434Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7104780Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7104820Z graph_break []
2025-12-04T09:39:16.7104917Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7104976Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7105072Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7105420Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7105458Z graph_break []
2025-12-04T09:39:16.7105535Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7105591Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7105713Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7106063Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7106099Z graph_break []
2025-12-04T09:39:16.7106172Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7106225Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7106322Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7106664Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7106703Z graph_break []
2025-12-04T09:39:16.7106804Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.7106852Z Traceback (most recent call last):
2025-12-04T09:39:16.7106981Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.7107023Z     self.common(
2025-12-04T09:39:16.7107112Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.7107157Z     return func(*args, **kwds)
2025-12-04T09:39:16.7107194Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.7107324Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 725, in check_model_gpu
2025-12-04T09:39:16.7107360Z     check_model(
2025-12-04T09:39:16.7107481Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.7107520Z     assert_equal_fn(
2025-12-04T09:39:16.7107665Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.7107724Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.7107770Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.7107932Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.7108006Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.7108058Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7108062Z 
2025-12-04T09:39:16.7108106Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7108203Z Greatest absolute difference: 0.58544921875 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7108295Z Greatest relative difference: 0.57568359375 at index (0, 1) (up to 0.001 allowed)
2025-12-04T09:39:16.7108297Z 
2025-12-04T09:39:16.7108344Z The failure occurred for item [2]
2025-12-04T09:39:16.7108347Z 
2025-12-04T09:39:16.7108443Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7108655Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7108657Z 
2025-12-04T09:39:16.7108742Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7108816Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7108870Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7109187Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7109284Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7109352Z graph_break []
2025-12-04T09:39:16.7109425Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7109483Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7109578Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7109931Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7109968Z graph_break []
2025-12-04T09:39:16.7110039Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7110097Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7110193Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7110583Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7110620Z graph_break []
2025-12-04T09:39:16.7110694Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7110749Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7110846Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7111194Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7111232Z graph_break []
2025-12-04T09:39:16.7111305Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7111362Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7111457Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7111810Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7111848Z graph_break []
2025-12-04T09:39:16.7111919Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7111976Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7112074Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7112460Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7112496Z graph_break []
2025-12-04T09:39:16.7112569Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7112622Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7112718Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7113062Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7113132Z graph_break []
2025-12-04T09:39:16.7113204Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7113261Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7113356Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7113706Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7113742Z graph_break []
2025-12-04T09:39:16.7113813Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7113870Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7113964Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7114316Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7114352Z graph_break []
2025-12-04T09:39:16.7114425Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7114480Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7114576Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7114923Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7114961Z graph_break []
2025-12-04T09:39:16.7115033Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7115090Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7115186Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7115535Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7115572Z graph_break []
2025-12-04T09:39:16.7115642Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7115698Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7115793Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7116172Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7116208Z graph_break []
2025-12-04T09:39:16.7116283Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7116335Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7116430Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7116777Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7116836Z graph_break []
2025-12-04T09:39:16.7116909Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7116967Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7117064Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7117416Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7117454Z graph_break []
2025-12-04T09:39:16.7117528Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7117583Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7117679Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7118033Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7118069Z graph_break []
2025-12-04T09:39:16.7118142Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7118195Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7118292Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7118641Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7118680Z graph_break []
2025-12-04T09:39:16.7118756Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7118813Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7118911Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7119267Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7119304Z graph_break []
2025-12-04T09:39:16.7119377Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7119435Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7119530Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7119907Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7119943Z graph_break []
2025-12-04T09:39:16.7120017Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7120072Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7120169Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7120546Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7120610Z graph_break []
2025-12-04T09:39:16.7120685Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7120740Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7120837Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7121188Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7121226Z graph_break []
2025-12-04T09:39:16.7121297Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7121354Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7121449Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7121799Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7121834Z graph_break []
2025-12-04T09:39:16.7121907Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7121960Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7122057Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7122401Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7122441Z graph_break []
2025-12-04T09:39:16.7122516Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7122570Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7122666Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7123008Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7123046Z graph_break []
2025-12-04T09:39:16.7123117Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7123177Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7123272Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7123647Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7123684Z graph_break []
2025-12-04T09:39:16.7123757Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7123812Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7123909Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7124257Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7124314Z graph_break []
2025-12-04T09:39:16.7124388Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7124443Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7124540Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7124886Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7124923Z graph_break []
2025-12-04T09:39:16.7124994Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7125051Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7125146Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7125495Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7125531Z graph_break []
2025-12-04T09:39:16.7125604Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7125659Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7125760Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7126110Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7126154Z graph_break []
2025-12-04T09:39:16.7126232Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7126289Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7126389Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7126737Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7126779Z graph_break []
2025-12-04T09:39:16.7126855Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7126914Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7127009Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7127387Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7127425Z graph_break []
2025-12-04T09:39:16.7127501Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7127556Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7127656Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7128000Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7128065Z graph_break []
2025-12-04T09:39:16.7128143Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7128200Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7128300Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7128650Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7128691Z graph_break []
2025-12-04T09:39:16.7128764Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7128825Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7128921Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7129278Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7129315Z graph_break []
2025-12-04T09:39:16.7129391Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7129447Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7129548Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7129901Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7129940Z graph_break []
2025-12-04T09:39:16.7130018Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7130074Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7130174Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7130554Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7130596Z graph_break []
2025-12-04T09:39:16.7130669Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7130729Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7130825Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7131206Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7131245Z graph_break []
2025-12-04T09:39:16.7131320Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7131376Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7131475Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7131829Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7131897Z graph_break []
2025-12-04T09:39:16.7131975Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7132031Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7132131Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7132479Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7132520Z graph_break []
2025-12-04T09:39:16.7132593Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7132654Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7132751Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7133109Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7133146Z graph_break []
2025-12-04T09:39:16.7133222Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7133279Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7133378Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7133735Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7133775Z graph_break []
2025-12-04T09:39:16.7133852Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7133909Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7134009Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7134354Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7134395Z graph_break []
2025-12-04T09:39:16.7134468Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7134526Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7134621Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7134993Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7135032Z graph_break []
2025-12-04T09:39:16.7135110Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7135167Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7135266Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7135619Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7135684Z graph_break []
2025-12-04T09:39:16.7135762Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7135818Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7135918Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7136266Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7136308Z graph_break []
2025-12-04T09:39:16.7136380Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7136442Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7136538Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7136898Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7136936Z graph_break []
2025-12-04T09:39:16.7137041Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.7137089Z Traceback (most recent call last):
2025-12-04T09:39:16.7137223Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.7137263Z     self.common(
2025-12-04T09:39:16.7137355Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.7137400Z     return func(*args, **kwds)
2025-12-04T09:39:16.7137443Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.7137576Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.7137615Z     check_model(
2025-12-04T09:39:16.7137741Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.7137781Z     assert_equal_fn(
2025-12-04T09:39:16.7137928Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.7137989Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.7138039Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.7138205Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.7138281Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.7138336Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7138338Z 
2025-12-04T09:39:16.7138386Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7138491Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7138621Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7138624Z 
2025-12-04T09:39:16.7138671Z The failure occurred for item [2]
2025-12-04T09:39:16.7138673Z 
2025-12-04T09:39:16.7138750Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7138962Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7138965Z 
2025-12-04T09:39:16.7139056Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7139129Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7139190Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7139515Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7139648Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7139690Z graph_break []
2025-12-04T09:39:16.7139763Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7139824Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7139920Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7140276Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7140315Z graph_break []
2025-12-04T09:39:16.7140393Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7140477Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7140580Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7140941Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7140983Z graph_break []
2025-12-04T09:39:16.7141058Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7141121Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7141223Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7141574Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7141615Z graph_break []
2025-12-04T09:39:16.7141688Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7141751Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7141847Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7142202Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7142240Z graph_break []
2025-12-04T09:39:16.7142317Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7142412Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7142513Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7142865Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7142909Z graph_break []
2025-12-04T09:39:16.7142982Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7143042Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7143142Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7143520Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7143563Z graph_break []
2025-12-04T09:39:16.7143636Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7143695Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7143793Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7144142Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7144181Z graph_break []
2025-12-04T09:39:16.7144257Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7144315Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7144416Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7144763Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7144803Z graph_break []
2025-12-04T09:39:16.7144876Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7144936Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7145035Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7145387Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7145428Z graph_break []
2025-12-04T09:39:16.7145503Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7145564Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7145660Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7146014Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7146052Z graph_break []
2025-12-04T09:39:16.7146131Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7146210Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7146314Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7146660Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7146702Z graph_break []
2025-12-04T09:39:16.7146779Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7146835Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7146934Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7147302Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7147343Z graph_break []
2025-12-04T09:39:16.7147416Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7147476Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7147574Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7147926Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7147965Z graph_break []
2025-12-04T09:39:16.7148041Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7148100Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7148202Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7148550Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7148593Z graph_break []
2025-12-04T09:39:16.7148671Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7148726Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7148826Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7149174Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7149216Z graph_break []
2025-12-04T09:39:16.7149289Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7149350Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7149447Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7149808Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7149848Z graph_break []
2025-12-04T09:39:16.7149924Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7150003Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7150104Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7150486Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7150530Z graph_break []
2025-12-04T09:39:16.7150607Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7150664Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7150764Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7151146Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7151188Z graph_break []
2025-12-04T09:39:16.7151263Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7151323Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7151420Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7151770Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7151809Z graph_break []
2025-12-04T09:39:16.7151885Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7151944Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7152043Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7152390Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7152432Z graph_break []
2025-12-04T09:39:16.7152508Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7152565Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7152665Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7153012Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7153055Z graph_break []
2025-12-04T09:39:16.7153129Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7153188Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7153285Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7153633Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7153672Z graph_break []
2025-12-04T09:39:16.7153749Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7153834Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7153934Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7154285Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7154327Z graph_break []
2025-12-04T09:39:16.7154402Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7154460Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7154560Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7154951Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7154991Z graph_break []
2025-12-04T09:39:16.7155064Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7155121Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7155219Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7155578Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7155615Z graph_break []
2025-12-04T09:39:16.7155697Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7155756Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7155857Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7156216Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7156253Z graph_break []
2025-12-04T09:39:16.7156330Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7156386Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7156487Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7156845Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7156887Z graph_break []
2025-12-04T09:39:16.7156962Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7157021Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7157117Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7157469Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7157507Z graph_break []
2025-12-04T09:39:16.7157587Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7157665Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7157767Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7158115Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7158154Z graph_break []
2025-12-04T09:39:16.7158233Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7158288Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7158388Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7158738Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7158804Z graph_break []
2025-12-04T09:39:16.7158878Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7158938Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7159036Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7159395Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7159433Z graph_break []
2025-12-04T09:39:16.7159512Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7159569Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7159670Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7160027Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7160064Z graph_break []
2025-12-04T09:39:16.7160141Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7160197Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7160298Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7160685Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7160728Z graph_break []
2025-12-04T09:39:16.7160801Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7160864Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7160960Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7161312Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7161351Z graph_break []
2025-12-04T09:39:16.7161432Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7161520Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7161619Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7161968Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7162007Z graph_break []
2025-12-04T09:39:16.7162085Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7162142Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7162244Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7162605Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7162680Z graph_break []
2025-12-04T09:39:16.7162754Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7162816Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7162912Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7163259Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7163297Z graph_break []
2025-12-04T09:39:16.7163376Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7163435Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7163536Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7163888Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7163925Z graph_break []
2025-12-04T09:39:16.7164002Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7164059Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7164160Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7164513Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7164556Z graph_break []
2025-12-04T09:39:16.7164631Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7164691Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7164790Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7165142Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7165180Z graph_break []
2025-12-04T09:39:16.7165258Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7165336Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7165436Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7165784Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7165823Z graph_break []
2025-12-04T09:39:16.7165899Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7165955Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7166056Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7166417Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7166479Z graph_break []
2025-12-04T09:39:16.7166553Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7166611Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7166709Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7167057Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7167094Z graph_break []
2025-12-04T09:39:16.7167171Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7167230Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7167327Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7167676Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7167713Z graph_break []
2025-12-04T09:39:16.7167786Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7167841Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7167940Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7170827Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7170878Z graph_break []
2025-12-04T09:39:16.7170984Z _ DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda __
2025-12-04T09:39:16.7171040Z Traceback (most recent call last):
2025-12-04T09:39:16.7171176Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 5770, in test_var_mean
2025-12-04T09:39:16.7171220Z     self.common(
2025-12-04T09:39:16.7171315Z   File "/opt/conda/envs/py_3.12/lib/python3.12/contextlib.py", line 81, in inner
2025-12-04T09:39:16.7171365Z     return func(*args, **kwds)
2025-12-04T09:39:16.7171404Z            ^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.7171540Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 692, in check_model_gpu
2025-12-04T09:39:16.7171583Z     check_model(
2025-12-04T09:39:16.7171754Z   File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 566, in check_model
2025-12-04T09:39:16.7171794Z     assert_equal_fn(
2025-12-04T09:39:16.7171946Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual
2025-12-04T09:39:16.7172008Z     return super().assertEqual(x, y, *args, **kwargs)
2025-12-04T09:39:16.7172061Z            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-12-04T09:39:16.7172234Z   File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 4284, in assertEqual
2025-12-04T09:39:16.7172308Z     raise error_metas.pop()[0].to_error(  # type: ignore[index]
2025-12-04T09:39:16.7172368Z AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7172371Z 
2025-12-04T09:39:16.7172419Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7172528Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7172664Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7172668Z 
2025-12-04T09:39:16.7172722Z The failure occurred for item [2]
2025-12-04T09:39:16.7172724Z 
2025-12-04T09:39:16.7172799Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7173023Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7173026Z 
2025-12-04T09:39:16.7173114Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7173195Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7173252Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7173586Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7173694Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7173732Z graph_break []
2025-12-04T09:39:16.7173809Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7173868Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7173972Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7174324Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7174365Z graph_break []
2025-12-04T09:39:16.7174442Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7174506Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7174605Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7174957Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7174995Z graph_break []
2025-12-04T09:39:16.7175074Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7175131Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7175232Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7175624Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7175665Z graph_break []
2025-12-04T09:39:16.7175744Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7175802Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7175904Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7176254Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7176296Z graph_break []
2025-12-04T09:39:16.7176370Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7176458Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7176556Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7176910Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7176947Z graph_break []
2025-12-04T09:39:16.7177026Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7177083Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7177185Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7177539Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7177580Z graph_break []
2025-12-04T09:39:16.7177661Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7177717Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7177817Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7178167Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7178208Z graph_break []
2025-12-04T09:39:16.7178284Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7178349Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7178448Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7178800Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7178835Z graph_break []
2025-12-04T09:39:16.7178910Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7178965Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7179068Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7179443Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7179483Z graph_break []
2025-12-04T09:39:16.7179563Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7179620Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7179723Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7180074Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7180116Z graph_break []
2025-12-04T09:39:16.7180189Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7180279Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7180376Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7180756Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7180795Z graph_break []
2025-12-04T09:39:16.7180871Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7180929Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7181030Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7181383Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7181423Z graph_break []
2025-12-04T09:39:16.7181503Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7181561Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7181663Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7182011Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7182052Z graph_break []
2025-12-04T09:39:16.7182126Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7182189Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7182288Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7182637Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7182675Z graph_break []
2025-12-04T09:39:16.7182752Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7182807Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7182908Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7183283Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7183324Z graph_break []
2025-12-04T09:39:16.7183401Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7183458Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7183562Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7183911Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7183952Z graph_break []
2025-12-04T09:39:16.7184027Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7184123Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7184223Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7184579Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7184617Z graph_break []
2025-12-04T09:39:16.7184695Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7184752Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7184855Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7185209Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7185249Z graph_break []
2025-12-04T09:39:16.7185325Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7185382Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7185485Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7185833Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7185877Z graph_break []
2025-12-04T09:39:16.7185953Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7186016Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7186115Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7186470Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7186507Z graph_break []
2025-12-04T09:39:16.7186585Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7186642Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7186742Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7187123Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7187163Z graph_break []
2025-12-04T09:39:16.7187242Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7187297Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7187398Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7187746Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7187789Z graph_break []
2025-12-04T09:39:16.7187863Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7187957Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7188056Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7188411Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7188450Z graph_break []
2025-12-04T09:39:16.7188530Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7188586Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7188690Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7189046Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7189085Z graph_break []
2025-12-04T09:39:16.7189164Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7189219Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7189320Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7189666Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7189707Z graph_break []
2025-12-04T09:39:16.7189780Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7189842Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7189939Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7190297Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7190336Z graph_break []
2025-12-04T09:39:16.7190450Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7190508Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7190613Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7191001Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7191041Z graph_break []
2025-12-04T09:39:16.7191116Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7191172Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7191270Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7191622Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7191667Z graph_break []
2025-12-04T09:39:16.7191741Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7191830Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7191928Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7192275Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7192313Z graph_break []
2025-12-04T09:39:16.7192387Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7192445Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7192543Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7192899Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7192943Z graph_break []
2025-12-04T09:39:16.7193020Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7193081Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7193177Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7193528Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7193568Z graph_break []
2025-12-04T09:39:16.7193644Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7193702Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7193804Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7194151Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7194191Z graph_break []
2025-12-04T09:39:16.7194264Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7194324Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7194422Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7194802Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7194847Z graph_break []
2025-12-04T09:39:16.7194920Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7194979Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7195076Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7195428Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7195465Z graph_break []
2025-12-04T09:39:16.7195539Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7195624Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7195725Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7196066Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7196110Z graph_break []
2025-12-04T09:39:16.7196184Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7196245Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7196346Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7196704Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7196747Z graph_break []
2025-12-04T09:39:16.7196820Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7196879Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7196979Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7197335Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7197372Z graph_break []
2025-12-04T09:39:16.7197448Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7197506Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7197610Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7197954Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7197996Z graph_break []
2025-12-04T09:39:16.7198070Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7198131Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7198232Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7198605Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7198649Z graph_break []
2025-12-04T09:39:16.7198721Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7198783Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7198880Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7199239Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7199277Z graph_break []
2025-12-04T09:39:16.7199353Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7199440Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7199542Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7199888Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7199929Z graph_break []
2025-12-04T09:39:16.7200002Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7200063Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7200163Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7200548Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7200591Z graph_break []
2025-12-04T09:39:16.7200663Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7200720Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7200817Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7201167Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7201205Z graph_break []
2025-12-04T09:39:16.7201282Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7201341Z stats [('calls_captured', 12), ('unique_graphs', 2)]
2025-12-04T09:39:16.7201445Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:39:16.7201793Z inductor [('triton_bundler_save_kernel', 112), ('benchmarking.InductorBenchmarker.benchmark', 12), ('benchmarking.InductorBenchmarker.benchmark_gpu', 12), ('async_compile_cache_miss', 8), ('async_compile_cache_hit', 4), ('fxgraph_cache_miss', 2), ('triton_bundler_save_static_autotuner', 2)]
2025-12-04T09:39:16.7201833Z graph_break []
2025-12-04T09:39:16.7201908Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7201963Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7202062Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7202435Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7202477Z graph_break []
2025-12-04T09:39:16.7202549Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:39:16.7202606Z stats [('calls_captured', 6), ('unique_graphs', 1)]
2025-12-04T09:39:16.7202703Z aot_autograd [('total', 1), ('autograd_cache_miss', 1), ('autograd_cache_saved', 1), ('ok', 1)]
2025-12-04T09:39:16.7203053Z inductor [('triton_bundler_save_kernel', 56), ('benchmarking.InductorBenchmarker.benchmark', 6), ('benchmarking.InductorBenchmarker.benchmark_gpu', 6), ('async_compile_cache_miss', 4), ('async_compile_cache_hit', 2), ('fxgraph_cache_miss', 1), ('triton_bundler_save_static_autotuner', 1)]
2025-12-04T09:39:16.7203091Z graph_break []
2025-12-04T09:39:16.7203371Z - generated xml file: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_torchinductor_dynamic_shapes/inductor.test_torchinductor_dynamic_shapes-1e4909ab018e7c6d.xml -
2025-12-04T09:39:16.7203466Z =========================== short test summary info ============================
2025-12-04T09:39:16.7203737Z FAILED [2.1502s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7203740Z 
2025-12-04T09:39:16.7203790Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7203894Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7204000Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7204002Z 
2025-12-04T09:39:16.7204049Z The failure occurred for item [2]
2025-12-04T09:39:16.7204051Z 
2025-12-04T09:39:16.7204129Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7204346Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7204351Z 
2025-12-04T09:39:16.7204443Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7204681Z FAILED [0.9008s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7204684Z 
2025-12-04T09:39:16.7204732Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7204832Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7204936Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7204938Z 
2025-12-04T09:39:16.7204986Z The failure occurred for item [2]
2025-12-04T09:39:16.7204989Z 
2025-12-04T09:39:16.7205063Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7205285Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7205287Z 
2025-12-04T09:39:16.7205371Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7205611Z FAILED [1.1713s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7205613Z 
2025-12-04T09:39:16.7205657Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7205756Z Greatest absolute difference: 0.58544921875 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7205849Z Greatest relative difference: 0.57568359375 at index (0, 1) (up to 0.001 allowed)
2025-12-04T09:39:16.7205851Z 
2025-12-04T09:39:16.7205899Z The failure occurred for item [2]
2025-12-04T09:39:16.7205901Z 
2025-12-04T09:39:16.7205976Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7206233Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7206236Z 
2025-12-04T09:39:16.7206323Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7206558Z FAILED [0.5407s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7206560Z 
2025-12-04T09:39:16.7206607Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7206708Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7206810Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7206812Z 
2025-12-04T09:39:16.7206856Z The failure occurred for item [2]
2025-12-04T09:39:16.7206883Z 
2025-12-04T09:39:16.7206960Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7207174Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7207176Z 
2025-12-04T09:39:16.7207264Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7207502Z FAILED [0.5404s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7207508Z 
2025-12-04T09:39:16.7207552Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7207654Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7207754Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7207757Z 
2025-12-04T09:39:16.7207808Z The failure occurred for item [2]
2025-12-04T09:39:16.7207810Z 
2025-12-04T09:39:16.7207885Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7208096Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7208099Z 
2025-12-04T09:39:16.7208183Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7208421Z FAILED [0.5656s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7208423Z 
2025-12-04T09:39:16.7208466Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7208569Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7208669Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7208675Z 
2025-12-04T09:39:16.7208718Z The failure occurred for item [2]
2025-12-04T09:39:16.7208720Z 
2025-12-04T09:39:16.7208798Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7209006Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7209008Z 
2025-12-04T09:39:16.7209095Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7209332Z FAILED [0.5276s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7209334Z 
2025-12-04T09:39:16.7209381Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7209479Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7209583Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7209585Z 
2025-12-04T09:39:16.7209651Z The failure occurred for item [2]
2025-12-04T09:39:16.7209653Z 
2025-12-04T09:39:16.7209730Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7209940Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7209942Z 
2025-12-04T09:39:16.7210026Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7210359Z FAILED [1.4937s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7210362Z 
2025-12-04T09:39:16.7210424Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7210525Z Greatest absolute difference: 0.58544921875 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7210650Z Greatest relative difference: 0.57568359375 at index (0, 1) (up to 0.001 allowed)
2025-12-04T09:39:16.7210654Z 
2025-12-04T09:39:16.7210702Z The failure occurred for item [2]
2025-12-04T09:39:16.7210704Z 
2025-12-04T09:39:16.7210776Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7210985Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7210987Z 
2025-12-04T09:39:16.7211070Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7211308Z FAILED [0.5544s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7211310Z 
2025-12-04T09:39:16.7211358Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7211460Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7211564Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7211566Z 
2025-12-04T09:39:16.7211611Z The failure occurred for item [2]
2025-12-04T09:39:16.7211613Z 
2025-12-04T09:39:16.7211690Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7211897Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7211899Z 
2025-12-04T09:39:16.7211987Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7212220Z FAILED [0.5585s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7212225Z 
2025-12-04T09:39:16.7212271Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7212372Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7212475Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7212478Z 
2025-12-04T09:39:16.7212526Z The failure occurred for item [2]
2025-12-04T09:39:16.7212528Z 
2025-12-04T09:39:16.7212601Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7212814Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7212816Z 
2025-12-04T09:39:16.7212901Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7213140Z FAILED [0.5515s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7213145Z 
2025-12-04T09:39:16.7213188Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7213321Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7213423Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7213425Z 
2025-12-04T09:39:16.7213473Z The failure occurred for item [2]
2025-12-04T09:39:16.7213475Z 
2025-12-04T09:39:16.7213549Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7213756Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7213758Z 
2025-12-04T09:39:16.7213846Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7214082Z FAILED [0.5407s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7214116Z 
2025-12-04T09:39:16.7214167Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7214265Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7214367Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7214369Z 
2025-12-04T09:39:16.7214414Z The failure occurred for item [2]
2025-12-04T09:39:16.7214416Z 
2025-12-04T09:39:16.7214492Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7214700Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7214705Z 
2025-12-04T09:39:16.7214789Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7215029Z FAILED [0.5468s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7215035Z 
2025-12-04T09:39:16.7215079Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7215179Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7215279Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7215281Z 
2025-12-04T09:39:16.7215327Z The failure occurred for item [2]
2025-12-04T09:39:16.7215329Z 
2025-12-04T09:39:16.7215402Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7215613Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7215615Z 
2025-12-04T09:39:16.7215699Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7215940Z FAILED [1.0617s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7215942Z 
2025-12-04T09:39:16.7215990Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7216084Z Greatest absolute difference: 0.58544921875 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7216179Z Greatest relative difference: 0.57568359375 at index (0, 1) (up to 0.001 allowed)
2025-12-04T09:39:16.7216181Z 
2025-12-04T09:39:16.7216226Z The failure occurred for item [2]
2025-12-04T09:39:16.7216228Z 
2025-12-04T09:39:16.7216304Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7216511Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7216512Z 
2025-12-04T09:39:16.7216599Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7216861Z FAILED [0.5359s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7216863Z 
2025-12-04T09:39:16.7216912Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7217011Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7217115Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7217117Z 
2025-12-04T09:39:16.7217164Z The failure occurred for item [2]
2025-12-04T09:39:16.7217166Z 
2025-12-04T09:39:16.7217240Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7217453Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7217455Z 
2025-12-04T09:39:16.7217560Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7217806Z FAILED [1.0560s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7217808Z 
2025-12-04T09:39:16.7217852Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7217948Z Greatest absolute difference: 0.58544921875 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7218042Z Greatest relative difference: 0.57568359375 at index (0, 1) (up to 0.001 allowed)
2025-12-04T09:39:16.7218043Z 
2025-12-04T09:39:16.7218091Z The failure occurred for item [2]
2025-12-04T09:39:16.7218093Z 
2025-12-04T09:39:16.7218165Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7218374Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7218378Z 
2025-12-04T09:39:16.7218467Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7218710Z FAILED [0.7166s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7218712Z 
2025-12-04T09:39:16.7218760Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7218858Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7218960Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7218962Z 
2025-12-04T09:39:16.7219008Z The failure occurred for item [2]
2025-12-04T09:39:16.7219009Z 
2025-12-04T09:39:16.7219087Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7219294Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7219299Z 
2025-12-04T09:39:16.7219387Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7219623Z FAILED [0.7142s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7219631Z 
2025-12-04T09:39:16.7219675Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7219777Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7219876Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7219878Z 
2025-12-04T09:39:16.7219927Z The failure occurred for item [2]
2025-12-04T09:39:16.7219929Z 
2025-12-04T09:39:16.7220000Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7220213Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7220238Z 
2025-12-04T09:39:16.7220323Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7220597Z FAILED [1.3510s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7220599Z 
2025-12-04T09:39:16.7220643Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7220740Z Greatest absolute difference: 0.58544921875 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7220832Z Greatest relative difference: 0.57568359375 at index (0, 1) (up to 0.001 allowed)
2025-12-04T09:39:16.7220838Z 
2025-12-04T09:39:16.7220884Z The failure occurred for item [2]
2025-12-04T09:39:16.7220885Z 
2025-12-04T09:39:16.7220962Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7221205Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7221207Z 
2025-12-04T09:39:16.7221295Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7221531Z FAILED [0.6883s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7221533Z 
2025-12-04T09:39:16.7221579Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7221677Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7221780Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7221782Z 
2025-12-04T09:39:16.7221825Z The failure occurred for item [2]
2025-12-04T09:39:16.7221827Z 
2025-12-04T09:39:16.7221903Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7222116Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7222118Z 
2025-12-04T09:39:16.7222202Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7222441Z FAILED [0.6755s] inductor/test_torchinductor_dynamic_shapes.py::DynamicShapesGPUTests::test_var_mean_tile_reduction_True_dynamic_shapes_cuda - AssertionError: Tensor-likes are not close!
2025-12-04T09:39:16.7222443Z 
2025-12-04T09:39:16.7222488Z Mismatched elements: 4 / 4 (100.0%)
2025-12-04T09:39:16.7222588Z Greatest absolute difference: 0.5851404666900635 at index (0, 3) (up to 1e-05 allowed)
2025-12-04T09:39:16.7222686Z Greatest relative difference: 0.5756681561470032 at index (0, 1) (up to 1.3e-06 allowed)
2025-12-04T09:39:16.7222688Z 
2025-12-04T09:39:16.7222735Z The failure occurred for item [2]
2025-12-04T09:39:16.7222739Z 
2025-12-04T09:39:16.7222811Z To execute this test, run the following from the base repo dir:
2025-12-04T09:39:16.7223024Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_var_mean_tile_reduction_True_dynamic_shapes_cuda
2025-12-04T09:39:16.7223026Z 
2025-12-04T09:39:16.7223111Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:39:16.7223180Z ======================== 21 failed, 29 passed in 52.93s ========================
2025-12-04T09:39:16.7223181Z 
2025-12-04T09:39:16.7223400Z FINISHED PRINTING LOG FILE of inductor/test_torchinductor_dynamic_shapes 5/5 (test/test-reports/inductor.test_torchinductor_dynamic_shapes_5.5_7bd540a7dc87d591_.log)
2025-12-04T09:39:16.7223402Z 
2025-12-04T09:39:16.7223544Z Finished inductor/test_torchinductor_dynamic_shapes 5/5 ... [2025-12-04 09:39:16.659329][2246373.793282962], took 1.01min
2025-12-04T09:39:16.7223786Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:39:16.7223900Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:39:16.7224002Z GITHUB_RUN_ID, GITHUB_RUN_ATTEMPT, or ARTIFACTS_FILE_SUFFIX not set, not uploading
2025-12-04T09:39:16.7224052Z Uploading artifacts took 0.00 seconds
2025-12-04T09:39:16.7224123Z inductor/test_torchinductor_dynamic_shapes 5/5 failed!
2025-12-04T09:39:16.7224242Z Running inductor/test_torchinductor_opinfo 3/10 ... [2025-12-04 09:39:16.660544][2246373.794502442]
2025-12-04T09:39:16.7224294Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:39:16.7224684Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_torchinductor_opinfo.py', '--shard-id=3', '--num-shards=10', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:39:16.660712]
2025-12-04T09:39:26.2217549Z 
2025-12-04T09:39:26.2218674Z inductor/test_torchinductor_opinfo 3/10 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_torchinductor_opinfo_3.10_62ed9eb698fbd229_.log
2025-12-04T09:39:26.2246024Z Running 100 items in this shard: test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_bilinear_cuda_float64, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_stft_cuda_float32
2025-12-04T09:39:26.2263050Z 
2025-12-04T09:39:26.2263184Z Finished inductor/test_torchinductor_opinfo 3/10 ... [2025-12-04 09:39:26.221627][2246383.355579622], took 0.16min
2025-12-04T09:39:26.2263606Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:39:26.2263962Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:39:26.2264198Z Running inductor/test_torchinductor_opinfo 9/10 ... [2025-12-04 09:39:26.223085][2246383.357042528]
2025-12-04T09:39:26.2264399Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:39:26.2264860Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_torchinductor_opinfo.py', '--shard-id=9', '--num-shards=10', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:39:26.223293]
2025-12-04T09:41:10.3576275Z 
2025-12-04T09:41:10.3577504Z inductor/test_torchinductor_opinfo 9/10 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_torchinductor_opinfo_9.10_7f99ad43d769148d_.log
2025-12-04T09:41:10.3618624Z Running 150 items in this shard: test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_adaptive_avg_pool2d_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_nn_functional_interpolate_linear_cuda_float16, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32, test/inductor/test_torchinductor_opinfo.py::TestInductorOpInfoCUDA::test_comprehensive_pca_lowrank_cuda_float32
2025-12-04T09:41:10.3645203Z 
2025-12-04T09:41:10.3645360Z Finished inductor/test_torchinductor_opinfo 9/10 ... [2025-12-04 09:41:10.357442][2246487.491392889], took 1.74min
2025-12-04T09:41:10.3645773Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:41:10.3646128Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:41:10.3646347Z Running inductor/test_cpu_repro 3/4 ... [2025-12-04 09:41:10.359009][2246487.492966463]
2025-12-04T09:41:10.3646531Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:41:10.3646977Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_cpu_repro.py', '--shard-id=3', '--num-shards=4', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:41:10.359195]
2025-12-04T09:41:17.3649813Z 
2025-12-04T09:41:17.3651132Z inductor/test_cpu_repro 3/4 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_cpu_repro_3.4_23f3abdb3149edbb_.log
2025-12-04T09:41:17.3651980Z Running 0 items in this shard:
2025-12-04T09:41:17.3652195Z 
2025-12-04T09:41:17.3652523Z Finished inductor/test_cpu_repro 3/4 ... [2025-12-04 09:41:17.364707][2246494.498660701], took 0.12min
2025-12-04T09:41:17.3657748Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:41:17.3660163Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:41:17.3660685Z Running dynamo/test_higher_order_ops 1/1 ... [2025-12-04 09:41:17.365915][2246494.499873241]
2025-12-04T09:41:17.3661034Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:41:17.3663401Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'dynamo/test_higher_order_ops.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:41:17.366076]
2025-12-04T09:41:23.4722980Z 
2025-12-04T09:41:23.4723930Z dynamo/test_higher_order_ops 1/1 was successful, full logs can be found in artifacts with path test/test-reports/dynamo.test_higher_order_ops_1.1_0b89d70e214bd61d_.log
2025-12-04T09:41:23.4724557Z Running 0 items in this shard:
2025-12-04T09:41:23.4724706Z 
2025-12-04T09:41:23.4724952Z Finished dynamo/test_higher_order_ops 1/1 ... [2025-12-04 09:41:23.471909][2246500.605860633], took 0.10min
2025-12-04T09:41:23.4727700Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:41:23.4733821Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:41:23.4739829Z Running inductor/test_custom_lowering 1/1 ... [2025-12-04 09:41:23.473605][2246500.607561854]
2025-12-04T09:41:23.4740135Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:41:23.4741001Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_custom_lowering.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:41:23.473835]
2025-12-04T09:41:28.7154085Z 
2025-12-04T09:41:28.7154766Z inductor/test_custom_lowering 1/1 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_custom_lowering_1.1_3f3adb571e57991c_.log
2025-12-04T09:41:28.7155131Z Running 0 items in this shard:
2025-12-04T09:41:28.7155216Z 
2025-12-04T09:41:28.7155350Z Finished inductor/test_custom_lowering 1/1 ... [2025-12-04 09:41:28.715095][2246505.849049574], took 0.09min
2025-12-04T09:41:28.7157392Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:41:28.7163099Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:41:28.7166355Z Running inductor/test_fused_attention 1/1 ... [2025-12-04 09:41:28.716506][2246505.850463591]
2025-12-04T09:41:28.7166807Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:41:28.7168668Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_fused_attention.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:41:28.716700]
2025-12-04T09:41:34.0549905Z 
2025-12-04T09:41:34.0550917Z inductor/test_fused_attention 1/1 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_fused_attention_1.1_a22eee43d5c522bb_.log
2025-12-04T09:41:34.0551548Z Running 0 items in this shard:
2025-12-04T09:41:34.0551686Z 
2025-12-04T09:41:34.0551900Z Finished inductor/test_fused_attention 1/1 ... [2025-12-04 09:41:34.054730][2246511.188682226], took 0.09min
2025-12-04T09:41:34.0556642Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:41:34.0561775Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:41:34.0564902Z Running inductor/test_smoke 1/1 ... [2025-12-04 09:41:34.056403][2246511.190360508]
2025-12-04T09:41:34.0565189Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:41:34.0567477Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_smoke.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:41:34.056628]
2025-12-04T09:41:39.2087408Z 
2025-12-04T09:41:39.2088362Z inductor/test_smoke 1/1 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_smoke_1.1_3c3fb90c1b38f715_.log
2025-12-04T09:41:39.2088864Z 
2025-12-04T09:41:39.2089102Z Finished inductor/test_smoke 1/1 ... [2025-12-04 09:41:39.208382][2246516.342335285], took 0.09min
2025-12-04T09:41:39.2091317Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:41:39.2099279Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:41:39.2100661Z Running inductor/test_flex_attention 1/4 ... [2025-12-04 09:41:39.209916][2246516.343874029]
2025-12-04T09:41:39.2101003Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:41:39.2103578Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_flex_attention.py', '--shard-id=1', '--num-shards=4', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:41:39.210120]
2025-12-04T09:45:15.3258099Z 
2025-12-04T09:45:15.3258709Z PRINTING LOG FILE of inductor/test_flex_attention 1/4 (test/test-reports/inductor.test_flex_attention_1.4_1061c3085781a0ce_.log)
2025-12-04T09:45:15.3268610Z Test results will be stored in test-reports/python-pytest/inductor.test_flex_attention/inductor.test_flex_attention-f823fef124b8972d.xml
2025-12-04T09:45:15.3268930Z ============================= test session starts ==============================
2025-12-04T09:45:15.3269181Z platform linux -- Python 3.12.5, pytest-7.3.2, pluggy-1.6.0 -- /opt/conda/envs/py_3.12/bin/python
2025-12-04T09:45:15.3269390Z cachedir: .pytest_cache
2025-12-04T09:45:15.3269632Z hypothesis profile 'pytorch_ci' -> database=None, max_examples=50, derandomize=True, suppress_health_check=[HealthCheck.too_slow]
2025-12-04T09:45:15.3270698Z rootdir: /var/lib/jenkins/pytorch
2025-12-04T09:45:15.3270827Z configfile: pytest.ini
2025-12-04T09:45:15.3271070Z plugins: hypothesis-6.56.4, cpp-2.3.0, flakefinder-1.1.0, rerunfailures-14.0, subtests-0.13.1, xdist-3.3.1, xdoctest-1.3.0, typeguard-4.3.0
2025-12-04T09:45:15.3271324Z collecting ... collected 763 items
2025-12-04T09:45:15.3271483Z stepcurrent: Cannot find last run test, not skipping
2025-12-04T09:45:15.3279473Z Running 50 items in this shard: test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda, test/inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda
2025-12-04T09:45:15.3286322Z 
2025-12-04T09:45:15.3286479Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda PASSED [6.1128s] [  2%]
2025-12-04T09:45:15.3286810Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.2083s] [  2%]
2025-12-04T09:45:15.3287138Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.8167s] [  2%]
2025-12-04T09:45:15.3287476Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.6221s] [  2%]
2025-12-04T09:45:15.3287795Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.7765s] [  2%]
2025-12-04T09:45:15.3288129Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.7979s] [  2%]
2025-12-04T09:45:15.3288452Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.9433s] [  2%]
2025-12-04T09:45:15.3288777Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.9359s] [  2%]
2025-12-04T09:45:15.3289099Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.7064s] [  2%]
2025-12-04T09:45:15.3289422Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.6811s] [  2%]
2025-12-04T09:45:15.3289746Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.8267s] [  2%]
2025-12-04T09:45:15.3290067Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.0667s] [  2%]
2025-12-04T09:45:15.3290514Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.1450s] [  2%]
2025-12-04T09:45:15.3290837Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.9231s] [  2%]
2025-12-04T09:45:15.3291160Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.9786s] [  2%]
2025-12-04T09:45:15.3291483Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.2159s] [  2%]
2025-12-04T09:45:15.3291802Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.3649s] [  2%]
2025-12-04T09:45:15.3292128Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.2806s] [  2%]
2025-12-04T09:45:15.3292448Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.4002s] [  2%]
2025-12-04T09:45:15.3292809Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.7822s] [  2%]
2025-12-04T09:45:15.3293128Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.7276s] [  2%]
2025-12-04T09:45:15.3293449Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.5690s] [  2%]
2025-12-04T09:45:15.3293768Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.1873s] [  2%]
2025-12-04T09:45:15.3294090Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.0301s] [  2%]
2025-12-04T09:45:15.3294422Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.8084s] [  2%]
2025-12-04T09:45:15.3294735Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.0913s] [  2%]
2025-12-04T09:45:15.3295057Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.1417s] [  2%]
2025-12-04T09:45:15.3295398Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.1571s] [  2%]
2025-12-04T09:45:15.3295721Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.4631s] [  2%]
2025-12-04T09:45:15.3296039Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.9384s] [  2%]
2025-12-04T09:45:15.3296361Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.3235s] [  2%]
2025-12-04T09:45:15.3296681Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.9420s] [  2%]
2025-12-04T09:45:15.3297006Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.3835s] [  2%]
2025-12-04T09:45:15.3297332Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.0163s] [  2%]
2025-12-04T09:45:15.3297660Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.0669s] [  2%]
2025-12-04T09:45:15.3297989Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.3283s] [  2%]
2025-12-04T09:45:15.3298322Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.0865s] [  2%]
2025-12-04T09:45:15.3298653Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.5002s] [  2%]
2025-12-04T09:45:15.3298982Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.1988s] [  2%]
2025-12-04T09:45:15.3299307Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.3243s] [  2%]
2025-12-04T09:45:15.3299634Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.0704s] [  2%]
2025-12-04T09:45:15.3300000Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.6904s] [  2%]
2025-12-04T09:45:15.3300327Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.1336s] [  2%]
2025-12-04T09:45:15.3300699Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [3.9265s] [  2%]
2025-12-04T09:45:15.3301030Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.3027s] [  2%]
2025-12-04T09:45:15.3301361Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.3690s] [  2%]
2025-12-04T09:45:15.3301680Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.2044s] [  2%]
2025-12-04T09:45:15.3302014Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.3931s] [  2%]
2025-12-04T09:45:15.3302380Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.3645s] [  2%]
2025-12-04T09:45:15.3302704Z inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda FAILED [4.5234s] [  2%]
2025-12-04T09:45:15.3302887Z 
2025-12-04T09:45:15.3302945Z =================================== FAILURES ===================================
2025-12-04T09:45:15.3303146Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.3303332Z Traceback (most recent call last):
2025-12-04T09:45:15.3303567Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.3303795Z     self.assertTrue(
2025-12-04T09:45:15.3303978Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.3304174Z     raise self.failureException(msg)
2025-12-04T09:45:15.3304399Z AssertionError: False is not true : Log file /tmp/tmpzpsl3_e9/flex_attention_configs.json was not created
2025-12-04T09:45:15.3304573Z 
2025-12-04T09:45:15.3304653Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.3304943Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.3305151Z 
2025-12-04T09:45:15.3305243Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.3305456Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.3305620Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.3305737Z unimplemented []
2025-12-04T09:45:15.3305866Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.3306555Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.3307293Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.3307482Z graph_break []
2025-12-04T09:45:15.3307619Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.3308241Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.3308822Z   current_size = base.storage().size()
2025-12-04T09:45:15.3309020Z Autotune Choices Stats:
2025-12-04T09:45:15.3309858Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.3310863Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3311214Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3311539Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3312381Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3313668Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3314940Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3316207Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3317530Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3318838Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3337277Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3338672Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3339938Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3341322Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3342122Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.3342371Z Autotune Choices Stats:
2025-12-04T09:45:15.3343255Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.3344300Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3344805Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3345314Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3346318Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3347713Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3349036Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3353575Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3355501Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3357277Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3359356Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3361147Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3363037Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3364529Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3365378Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.3365677Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.3365858Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.3365979Z unimplemented []
2025-12-04T09:45:15.3366126Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.3366349Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.3367140Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.3367846Z graph_break []
2025-12-04T09:45:15.3367992Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.3368853Z Autotune Choices Stats:
2025-12-04T09:45:15.3369809Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.3370849Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3371167Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3371514Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3372411Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3373797Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3375056Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3376312Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3377573Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3378875Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3380188Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3381516Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3382825Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3384100Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3384891Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.3385114Z Autotune Choices Stats:
2025-12-04T09:45:15.3385983Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.3387002Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3387439Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3387943Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3389003Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3390298Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3391706Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3393068Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3394369Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3395743Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3397036Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3398367Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3399835Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3401277Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3402097Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.3402371Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.3402561Z Traceback (most recent call last):
2025-12-04T09:45:15.3402826Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.3403068Z     self.assertTrue(
2025-12-04T09:45:15.3403269Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.3403466Z     raise self.failureException(msg)
2025-12-04T09:45:15.3403702Z AssertionError: False is not true : Log file /tmp/tmp_c38ejjw/flex_attention_configs.json was not created
2025-12-04T09:45:15.3403870Z 
2025-12-04T09:45:15.3403961Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.3404261Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.3404465Z 
2025-12-04T09:45:15.3404572Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.3404788Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.3404956Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.3405074Z unimplemented []
2025-12-04T09:45:15.3405207Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.3405902Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.3406622Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.3406806Z graph_break []
2025-12-04T09:45:15.3406987Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.3407605Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.3408189Z   current_size = base.storage().size()
2025-12-04T09:45:15.3408333Z Autotune Choices Stats:
2025-12-04T09:45:15.3409176Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.3410145Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3410520Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3410855Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3411692Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3412957Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3414231Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3415487Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3416782Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3418034Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3419338Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3420675Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3421948Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3423213Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3424004Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.3424233Z Autotune Choices Stats:
2025-12-04T09:45:15.3425136Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.3426153Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3426592Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3427086Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3428073Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3429390Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3430741Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3432049Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3433358Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3434716Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3436030Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3437346Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3438626Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3439918Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3440776Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.3441020Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.3441181Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.3441293Z unimplemented []
2025-12-04T09:45:15.3441416Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.3441617Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.3442345Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.3442990Z graph_break []
2025-12-04T09:45:15.3443122Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.3443278Z Autotune Choices Stats:
2025-12-04T09:45:15.3444119Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.3445211Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3445549Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3446021Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3446896Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3448186Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3449539Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3450908Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3452242Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3453576Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3454879Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3456236Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3457558Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3458883Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3459742Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.3460018Z Autotune Choices Stats:
2025-12-04T09:45:15.3460947Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.3462025Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3462547Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3463136Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3464142Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3465501Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3466873Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3468434Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3469863Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3471347Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3472748Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3474110Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3475531Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3476921Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3524402Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.3524810Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.3525166Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.3525393Z unimplemented []
2025-12-04T09:45:15.3525612Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.3525883Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.3526719Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.3527474Z graph_break []
2025-12-04T09:45:15.3527708Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.3527923Z Autotune Choices Stats:
2025-12-04T09:45:15.3528917Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.3530048Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3530473Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3530885Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3531787Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3533260Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3534654Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3535979Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3537318Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3538654Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3540019Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3541595Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3542958Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3544514Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3545360Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.3545630Z Autotune Choices Stats:
2025-12-04T09:45:15.3546561Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.3547701Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3548240Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3548815Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3550110Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3551593Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3553004Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3554388Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3555766Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3557180Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3558582Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3560039Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3561477Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3562872Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3563773Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.3564121Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.3564369Z Traceback (most recent call last):
2025-12-04T09:45:15.3564694Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.3565003Z     self.assertTrue(
2025-12-04T09:45:15.3565262Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.3565506Z     raise self.failureException(msg)
2025-12-04T09:45:15.3565839Z AssertionError: False is not true : Log file /tmp/tmpb8up9rbc/flex_attention_configs.json was not created
2025-12-04T09:45:15.3566057Z 
2025-12-04T09:45:15.3566249Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.3566624Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.3566860Z 
2025-12-04T09:45:15.3567055Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.3567302Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.3567545Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.3567699Z unimplemented []
2025-12-04T09:45:15.3567927Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.3568690Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.3569522Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.3569801Z graph_break []
2025-12-04T09:45:15.3570230Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.3571022Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.3571660Z   current_size = base.storage().size()
2025-12-04T09:45:15.3571863Z Autotune Choices Stats:
2025-12-04T09:45:15.3572732Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.3573773Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3574143Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3574531Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3575432Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3576771Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3578180Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3579506Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3580958Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3582302Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3583690Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3585085Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3586452Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3587804Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3588816Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.3589065Z Autotune Choices Stats:
2025-12-04T09:45:15.3589973Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.3591139Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3591657Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3592246Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3593347Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3594719Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3596115Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3597719Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3599138Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3600627Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3602018Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3603431Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3604843Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3606203Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3607099Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.3607397Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.3607627Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.3607833Z unimplemented []
2025-12-04T09:45:15.3608032Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.3608302Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.3609138Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.3609865Z graph_break []
2025-12-04T09:45:15.3610040Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.3610267Z Autotune Choices Stats:
2025-12-04T09:45:15.3611224Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.3612243Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3612731Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3613218Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3614126Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3615511Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3616830Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3618151Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3619458Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3620884Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3622231Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3623691Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3625079Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3626399Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3627271Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.3627604Z Autotune Choices Stats:
2025-12-04T09:45:15.3628519Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.3629611Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3630167Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3630786Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3631817Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3633355Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3634753Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3636179Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3637575Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3638977Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3640556Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3641937Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3643395Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3644854Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3645721Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.3646062Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.3646334Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.3646494Z unimplemented []
2025-12-04T09:45:15.3646714Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.3646958Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.3647806Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.3648508Z graph_break []
2025-12-04T09:45:15.3648721Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.3648953Z Autotune Choices Stats:
2025-12-04T09:45:15.3650076Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.3651138Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3651515Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3651952Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3652874Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3654301Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3655663Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3657039Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3658364Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3659689Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3661166Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3662505Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3663878Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3665213Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3666073Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.3666345Z Autotune Choices Stats:
2025-12-04T09:45:15.3667280Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.3668519Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3669068Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3669664Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3670797Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3754328Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3755954Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3757553Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3759389Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3761057Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3762576Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3764130Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3765641Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3767028Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3767889Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.3768188Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.3768393Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.3768755Z unimplemented []
2025-12-04T09:45:15.3768926Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.3769186Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.3769951Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.3770695Z graph_break []
2025-12-04T09:45:15.3770896Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.3771097Z Autotune Choices Stats:
2025-12-04T09:45:15.3771963Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.3772942Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3773274Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3773675Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3774617Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3781162Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3782574Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3783879Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3785295Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3786797Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3788177Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3789480Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3790830Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3792127Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3792908Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.3793122Z Autotune Choices Stats:
2025-12-04T09:45:15.3793972Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.3795019Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3795453Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3795950Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3797003Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3798355Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3799698Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3801483Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3802816Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3804191Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3805523Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3806849Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3808229Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3809627Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3811024Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.3811317Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.3811593Z Traceback (most recent call last):
2025-12-04T09:45:15.3811928Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.3812259Z     self.assertTrue(
2025-12-04T09:45:15.3812502Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.3812744Z     raise self.failureException(msg)
2025-12-04T09:45:15.3813070Z AssertionError: False is not true : Log file /tmp/tmpz0pv_51w/flex_attention_configs.json was not created
2025-12-04T09:45:15.3813304Z 
2025-12-04T09:45:15.3813427Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.3813792Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.3814040Z 
2025-12-04T09:45:15.3814178Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.3814461Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.3814696Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.3814875Z unimplemented []
2025-12-04T09:45:15.3815086Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.3815866Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.3816653Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.3816889Z graph_break []
2025-12-04T09:45:15.3817093Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.3817768Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.3818443Z   current_size = base.storage().size()
2025-12-04T09:45:15.3818605Z Autotune Choices Stats:
2025-12-04T09:45:15.3819536Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.3820615Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3820992Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3821439Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3822335Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3823688Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3825057Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3826434Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3827759Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3829191Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3830626Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3832128Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3833437Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3834882Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3835790Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.3836091Z Autotune Choices Stats:
2025-12-04T09:45:15.3837027Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.3838124Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3838640Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3839213Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3840392Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3841846Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3843233Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3844761Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3846230Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3847606Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3849013Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3850463Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3851893Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3853290Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3854320Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.3854660Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.3854860Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.3855064Z unimplemented []
2025-12-04T09:45:15.3855291Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.3855578Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.3856394Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.3857096Z graph_break []
2025-12-04T09:45:15.3857319Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.3857580Z Autotune Choices Stats:
2025-12-04T09:45:15.3858590Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.3859613Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3859970Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3860365Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3861354Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3862727Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3864125Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3865444Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3866833Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3868319Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3869672Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3871085Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3872455Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3873850Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3874698Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.3875014Z Autotune Choices Stats:
2025-12-04T09:45:15.3875939Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.3877059Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3877684Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3878245Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3879333Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3880773Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3882236Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3883668Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3885093Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3886494Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3887885Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3889412Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3890820Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3892247Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3893131Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.3893438Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.3893700Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.3893858Z unimplemented []
2025-12-04T09:45:15.3894052Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.3894361Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.3895250Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.3895959Z graph_break []
2025-12-04T09:45:15.3896159Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.3896383Z Autotune Choices Stats:
2025-12-04T09:45:15.3897253Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.3898236Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3898571Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3899011Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3899892Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3901317Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3902924Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3904279Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3905589Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3906935Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3908247Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3909634Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3911083Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3912432Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3913307Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.3913622Z Autotune Choices Stats:
2025-12-04T09:45:15.3914536Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.3915652Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.3916151Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.3916715Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.3917789Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3919218Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3920627Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3922105Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3923835Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3925228Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3926635Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.3928006Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3929440Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3930871Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.3931802Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.3932072Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.3932342Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.3932515Z unimplemented []
2025-12-04T09:45:15.3932706Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.3933009Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.3933828Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.3934556Z graph_break []
2025-12-04T09:45:15.3934756Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.3935016Z Autotune Choices Stats:
2025-12-04T09:45:15.3935911Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.4009367Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4009678Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4010002Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4010931Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4012213Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4013456Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4014757Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4016012Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4017264Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4018557Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4019835Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4021144Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4022420Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4023239Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.4023451Z Autotune Choices Stats:
2025-12-04T09:45:15.4024278Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.4025301Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4025736Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4026223Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4027178Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4028474Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4029837Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4031222Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4032562Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4033861Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4035173Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4036482Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4037803Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4039110Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4039909Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.4040181Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4040367Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4040513Z unimplemented []
2025-12-04T09:45:15.4040642Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4040844Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4041587Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4042351Z graph_break []
2025-12-04T09:45:15.4042495Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4042655Z Autotune Choices Stats:
2025-12-04T09:45:15.4043469Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.4044387Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4044684Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4045004Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4045833Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4047165Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4048425Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4049671Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4050998Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4052251Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4053520Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4054780Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4056167Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4057477Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4058247Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.4058457Z Autotune Choices Stats:
2025-12-04T09:45:15.4059313Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.4060320Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4060777Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4061261Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4062204Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4063495Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4064810Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4066094Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4067373Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4068763Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4070051Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4071666Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4072975Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4074290Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4075088Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.4075352Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.4075535Z Traceback (most recent call last):
2025-12-04T09:45:15.4075773Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.4075999Z     self.assertTrue(
2025-12-04T09:45:15.4076171Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.4076359Z     raise self.failureException(msg)
2025-12-04T09:45:15.4076571Z AssertionError: False is not true : Log file /tmp/tmp7_woya4z/flex_attention_configs.json was not created
2025-12-04T09:45:15.4076770Z 
2025-12-04T09:45:15.4076854Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.4077132Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.4077333Z 
2025-12-04T09:45:15.4077428Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.4077632Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4077792Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4077904Z unimplemented []
2025-12-04T09:45:15.4078028Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4078712Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.4079428Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4079606Z graph_break []
2025-12-04T09:45:15.4079738Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4080354Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.4080978Z   current_size = base.storage().size()
2025-12-04T09:45:15.4081106Z Autotune Choices Stats:
2025-12-04T09:45:15.4081922Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.4082823Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4083106Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4083458Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4084278Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4085529Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4086795Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4088036Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4089282Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4090583Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4091833Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4093115Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4094363Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4095634Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4096403Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.4096612Z Autotune Choices Stats:
2025-12-04T09:45:15.4097439Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.4098443Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4098870Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4099353Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4100305Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4101667Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4102947Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4104256Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4105539Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4106841Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4108135Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4109417Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4110774Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4112057Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4112872Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.4113118Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4113277Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4113390Z unimplemented []
2025-12-04T09:45:15.4113514Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4113712Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4114427Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4115071Z graph_break []
2025-12-04T09:45:15.4115202Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4115359Z Autotune Choices Stats:
2025-12-04T09:45:15.4116182Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.4117095Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4117379Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4117696Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4118547Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4119791Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4121064Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4122342Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4123576Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4124822Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4126078Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4127334Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4128595Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4129832Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4130664Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.4130872Z Autotune Choices Stats:
2025-12-04T09:45:15.4131710Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.4132727Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4133150Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4133647Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4134597Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4135891Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4137227Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4138508Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4139821Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4141149Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4142447Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4143737Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4145050Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4146331Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4147120Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.4147363Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4147552Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4147665Z unimplemented []
2025-12-04T09:45:15.4147787Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4147986Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4148704Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4149346Z graph_break []
2025-12-04T09:45:15.4149479Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4149637Z Autotune Choices Stats:
2025-12-04T09:45:15.4150555Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.4151454Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4151734Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4152054Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4152883Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4154166Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4155411Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4156755Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4158034Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4159320Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4160607Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4161865Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4163148Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4164399Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4165170Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.4165378Z Autotune Choices Stats:
2025-12-04T09:45:15.4166205Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.4167243Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4167670Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4168163Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4169120Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4170447Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4171737Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4173064Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4174351Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4175685Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4176982Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4178273Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4179575Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4180932Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4181729Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.4181976Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4182136Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4182251Z unimplemented []
2025-12-04T09:45:15.4182374Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4182575Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4183295Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.4183984Z graph_break []
2025-12-04T09:45:15.4184115Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4184271Z Autotune Choices Stats:
2025-12-04T09:45:15.4185088Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.4186002Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4186283Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4186599Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4187419Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4188672Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4189952Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4191244Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4192489Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4193778Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4195042Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4196303Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4197566Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4198848Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4199623Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.4199832Z Autotune Choices Stats:
2025-12-04T09:45:15.4200699Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.4201737Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4202161Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4202640Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4203610Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4204908Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4206193Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4207511Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4208808Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4210104Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4211465Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4212768Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4214079Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4215373Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4216169Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.4216411Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4216568Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4216683Z unimplemented []
2025-12-04T09:45:15.4216801Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4217041Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4217751Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4218391Z graph_break []
2025-12-04T09:45:15.4218522Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4218680Z Autotune Choices Stats:
2025-12-04T09:45:15.4219513Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.4220450Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4220728Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4221040Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4221859Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4223102Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4224351Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4225632Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4226871Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4228131Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4229409Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4230704Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4231948Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4233208Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4233977Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.4234187Z Autotune Choices Stats:
2025-12-04T09:45:15.4235051Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.4236060Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4236483Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4236997Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4237952Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4239253Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4240592Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4241876Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4243194Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4244488Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4245786Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4247097Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4248393Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4249676Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4250502Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.4250752Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4250912Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4251028Z unimplemented []
2025-12-04T09:45:15.4251149Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4251349Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4252099Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.4252747Z graph_break []
2025-12-04T09:45:15.4252877Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4253035Z Autotune Choices Stats:
2025-12-04T09:45:15.4253850Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.4254785Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4255068Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4255388Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4256214Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4257470Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4258737Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4259992Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4261328Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4262578Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4263854Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4265103Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4266359Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4267605Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4268385Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.4268595Z Autotune Choices Stats:
2025-12-04T09:45:15.4269421Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.4270575Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4271001Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4271482Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4272451Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4273774Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4275057Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4276346Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4277639Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4278956Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4280230Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4281547Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4282884Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4284191Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4284985Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.4285247Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.4285429Z Traceback (most recent call last):
2025-12-04T09:45:15.4285667Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.4285897Z     self.assertTrue(
2025-12-04T09:45:15.4286068Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.4286259Z     raise self.failureException(msg)
2025-12-04T09:45:15.4286473Z AssertionError: False is not true : Log file /tmp/tmplrzjtur0/flex_attention_configs.json was not created
2025-12-04T09:45:15.4286635Z 
2025-12-04T09:45:15.4286717Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.4286995Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.4287194Z 
2025-12-04T09:45:15.4287289Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.4287489Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4287649Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4287763Z unimplemented []
2025-12-04T09:45:15.4287921Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4288602Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.4289310Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4289487Z graph_break []
2025-12-04T09:45:15.4289619Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4290256Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.4290863Z   current_size = base.storage().size()
2025-12-04T09:45:15.4290988Z Autotune Choices Stats:
2025-12-04T09:45:15.4291804Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.4292714Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4292998Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4293314Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4294128Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4295388Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4296671Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4297915Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4299167Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4300474Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4301719Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4302964Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4304234Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4305578Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4306351Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.4306557Z Autotune Choices Stats:
2025-12-04T09:45:15.4307388Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.4308433Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4308858Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4309340Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4310303Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4311653Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4312960Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4314277Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4315552Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4316851Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4318164Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4319466Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4320789Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4322068Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4322858Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.4323100Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4323260Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4323376Z unimplemented []
2025-12-04T09:45:15.4323496Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4323730Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4324452Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4325092Z graph_break []
2025-12-04T09:45:15.4325224Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4325380Z Autotune Choices Stats:
2025-12-04T09:45:15.4326326Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.4327226Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4327506Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4327822Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4328644Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4329902Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4331185Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4332459Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4333702Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4334945Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4336212Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4337473Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4338728Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4339969Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4340781Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.4340989Z Autotune Choices Stats:
2025-12-04T09:45:15.4341846Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.4342850Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4343270Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4343778Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4344739Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4346033Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4347313Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4348603Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4349928Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4351256Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4352542Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4353846Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4355138Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4356430Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4357223Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.4357467Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4357626Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4357738Z unimplemented []
2025-12-04T09:45:15.4357860Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4358060Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4358803Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4359440Z graph_break []
2025-12-04T09:45:15.4359572Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4359728Z Autotune Choices Stats:
2025-12-04T09:45:15.4360574Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.4360734Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4360854Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4361017Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4361637Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4362242Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4362850Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4363473Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4364107Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4364715Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4365340Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4365952Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4366563Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4367168Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4367302Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.4367349Z Autotune Choices Stats:
2025-12-04T09:45:15.4368124Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.4368371Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4368543Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4368823Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4369456Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4370104Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4370770Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4371399Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4372027Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4372700Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4373327Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4373972Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4374626Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4375253Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4375384Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.4375464Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4375507Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4375550Z unimplemented []
2025-12-04T09:45:15.4375611Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4375716Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4376295Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.4376333Z graph_break []
2025-12-04T09:45:15.4376410Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4376452Z Autotune Choices Stats:
2025-12-04T09:45:15.4377218Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.4377348Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4377469Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4377633Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4378266Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4378875Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4379506Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4380110Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4380753Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4381388Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4381996Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4382624Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4383251Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4383862Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4383995Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.4384036Z Autotune Choices Stats:
2025-12-04T09:45:15.4384805Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.4385024Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4385194Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4385501Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4386133Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4386760Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4387405Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4388037Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4388669Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4389296Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4389939Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4390623Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4391278Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4391904Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4392039Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.4392114Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4392162Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4392200Z unimplemented []
2025-12-04T09:45:15.4392264Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4392365Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4392941Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4392984Z graph_break []
2025-12-04T09:45:15.4393058Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4393101Z Autotune Choices Stats:
2025-12-04T09:45:15.4393848Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.4394008Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4394124Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4394291Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4394915Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4395539Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4396148Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4396756Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4397362Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4397963Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4398594Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4399201Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4399834Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4400474Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4400609Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.4400654Z Autotune Choices Stats:
2025-12-04T09:45:15.4401422Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.4401648Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4401815Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4402096Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4402758Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4403386Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4404041Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4404662Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4405295Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4405923Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4406547Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4407215Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4407848Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4408496Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4408625Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.4408703Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4408747Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4408789Z unimplemented []
2025-12-04T09:45:15.4408851Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4408957Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4409533Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.4409575Z graph_break []
2025-12-04T09:45:15.4409652Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4409692Z Autotune Choices Stats:
2025-12-04T09:45:15.4410498Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.4410628Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4410747Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4410909Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4411750Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4412363Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4412997Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4413602Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4414234Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4414860Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4415473Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4416094Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4416704Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4417331Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4417460Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.4417506Z Autotune Choices Stats:
2025-12-04T09:45:15.4418274Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.4418493Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4418662Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4418949Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4419587Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4420244Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4420909Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4421567Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4422201Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4422832Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4423463Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4424088Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4424737Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4425362Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4425512Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.4425587Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4425633Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4425670Z unimplemented []
2025-12-04T09:45:15.4425735Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4425835Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4426423Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4426462Z graph_break []
2025-12-04T09:45:15.4426540Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4426583Z Autotune Choices Stats:
2025-12-04T09:45:15.4427338Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.4427472Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4427589Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4427754Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4470785Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4471686Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4472295Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4472967Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4473565Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4474163Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4474768Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4475375Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4476004Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4476603Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4476759Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.4476802Z Autotune Choices Stats:
2025-12-04T09:45:15.4477569Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.4477797Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4477968Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4478247Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4478879Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4479507Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4480147Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4480826Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4481482Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4482103Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4482723Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4483351Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4483979Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4484627Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4484762Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.4484859Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.4484909Z Traceback (most recent call last):
2025-12-04T09:45:15.4485067Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.4485138Z     self.assertTrue(
2025-12-04T09:45:15.4485249Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.4485298Z     raise self.failureException(msg)
2025-12-04T09:45:15.4485428Z AssertionError: False is not true : Log file /tmp/tmpdyu0elhy/flex_attention_configs.json was not created
2025-12-04T09:45:15.4485434Z 
2025-12-04T09:45:15.4485513Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.4485680Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.4485683Z 
2025-12-04T09:45:15.4485772Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.4485853Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4485898Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4485939Z unimplemented []
2025-12-04T09:45:15.4486004Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4486590Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.4486692Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4486729Z graph_break []
2025-12-04T09:45:15.4486806Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4487303Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.4487357Z   current_size = base.storage().size()
2025-12-04T09:45:15.4487398Z Autotune Choices Stats:
2025-12-04T09:45:15.4488153Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.4488308Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4488426Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4488590Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4489199Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4489821Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4490459Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4491060Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4491658Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4492257Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4492883Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4493485Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4494109Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4494709Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4494842Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.4494882Z Autotune Choices Stats:
2025-12-04T09:45:15.4495637Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.4495860Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4496027Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4496304Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4496995Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4497616Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4498260Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4498880Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4499503Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4500126Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4500771Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4501428Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4502050Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4502694Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4502824Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.4502900Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4502942Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4502984Z unimplemented []
2025-12-04T09:45:15.4503047Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4503151Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4503723Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4503762Z graph_break []
2025-12-04T09:45:15.4503834Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4503875Z Autotune Choices Stats:
2025-12-04T09:45:15.4504607Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.4504738Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4504854Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4505014Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4505648Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4506250Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4506870Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4507471Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4508079Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4508682Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4509282Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4509904Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4510540Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4511165Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4511293Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.4511335Z Autotune Choices Stats:
2025-12-04T09:45:15.4512092Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.4512316Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4512484Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4512762Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4513388Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4514035Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4514656Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4515292Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4515909Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4516533Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4517149Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4517769Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4518417Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4519033Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4519181Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.4519257Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4519301Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4519339Z unimplemented []
2025-12-04T09:45:15.4519401Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4519500Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4520074Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4520112Z graph_break []
2025-12-04T09:45:15.4520187Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4520226Z Autotune Choices Stats:
2025-12-04T09:45:15.4520995Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.4521128Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4521243Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4521402Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4522008Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4522638Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4523240Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4523866Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4524464Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4525072Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4525675Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4526276Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4526901Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4527506Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4527657Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.4527697Z Autotune Choices Stats:
2025-12-04T09:45:15.4528457Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.4528680Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4528844Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4529124Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4529758Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4530384Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4531061Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4531684Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4532341Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4532966Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4533591Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4534217Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4534840Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4535480Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4535610Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.4535685Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4535730Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4535769Z unimplemented []
2025-12-04T09:45:15.4535830Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4535960Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4536534Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.4536573Z graph_break []
2025-12-04T09:45:15.4536646Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4536686Z Autotune Choices Stats:
2025-12-04T09:45:15.4537419Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.4537549Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4537665Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4537822Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4538439Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4539058Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4539660Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4540257Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4540905Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4541513Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4542119Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4542716Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4543316Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4543952Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4544080Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.4544121Z Autotune Choices Stats:
2025-12-04T09:45:15.4544884Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.4545127Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4545295Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4545574Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4546208Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4546831Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4547452Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4550748Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4551374Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4552061Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4552684Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4553312Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4553935Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4554556Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4554764Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.4554843Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4554886Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4554924Z unimplemented []
2025-12-04T09:45:15.4554984Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4555086Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4555662Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4555716Z graph_break []
2025-12-04T09:45:15.4555789Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4555831Z Autotune Choices Stats:
2025-12-04T09:45:15.4556572Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.4556704Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4556820Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4556982Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4557601Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4558211Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4558834Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4559453Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4560049Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4560698Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4561306Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4561912Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4562518Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4563122Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4563291Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.4563335Z Autotune Choices Stats:
2025-12-04T09:45:15.4564097Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.4564329Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4564496Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4564773Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4565406Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4566035Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4566660Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4567280Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4567940Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4568566Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4569198Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4569822Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4570496Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4571113Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4571244Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.4571315Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4571359Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4571395Z unimplemented []
2025-12-04T09:45:15.4571456Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4571581Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4572167Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.4572203Z graph_break []
2025-12-04T09:45:15.4572279Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4572318Z Autotune Choices Stats:
2025-12-04T09:45:15.4573054Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.4573197Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4573311Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4573474Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4574085Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4574690Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4575287Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4575904Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4576511Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4577113Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4577721Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4578326Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4578928Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4579532Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4579661Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.4579702Z Autotune Choices Stats:
2025-12-04T09:45:15.4580512Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.4580745Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4580912Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4581208Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4581839Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4582459Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4583077Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4583701Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4584356Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4584988Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4585610Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4586243Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4586868Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4587492Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4587624Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.4587699Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4587741Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4587779Z unimplemented []
2025-12-04T09:45:15.4587838Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4587939Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4588528Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4588580Z graph_break []
2025-12-04T09:45:15.4588652Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4588695Z Autotune Choices Stats:
2025-12-04T09:45:15.4589439Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.4589580Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4589698Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4589857Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4590502Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4591113Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4591732Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4592340Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4592969Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4593587Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4594194Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4594813Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4595418Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4596037Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4596169Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.4596211Z Autotune Choices Stats:
2025-12-04T09:45:15.4596970Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.4597213Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4597403Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4597682Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4598312Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4598955Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4599596Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4600218Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4600878Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4601536Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4602174Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4602803Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4603453Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4604081Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4604214Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.4604288Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4604330Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4604368Z unimplemented []
2025-12-04T09:45:15.4604430Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4604531Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4605109Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.4605145Z graph_break []
2025-12-04T09:45:15.4605222Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4605261Z Autotune Choices Stats:
2025-12-04T09:45:15.4606028Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.4606172Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4606285Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4606449Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4607073Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4607680Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4608286Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4608891Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4609494Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4610124Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4610770Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4611379Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4612013Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4612621Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4612754Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.4612794Z Autotune Choices Stats:
2025-12-04T09:45:15.4613557Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.4613781Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4613948Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4614261Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4614918Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4615547Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4616192Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4616817Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4617453Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4618089Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4618739Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4619381Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4620010Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4620680Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4620815Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.4620908Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.4620956Z Traceback (most recent call last):
2025-12-04T09:45:15.4621209Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.4621449Z     self.assertTrue(
2025-12-04T09:45:15.4621627Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.4621825Z     raise self.failureException(msg)
2025-12-04T09:45:15.4622048Z AssertionError: False is not true : Log file /tmp/tmp4919na5_/flex_attention_configs.json was not created
2025-12-04T09:45:15.4622222Z 
2025-12-04T09:45:15.4622306Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.4622584Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.4622789Z 
2025-12-04T09:45:15.4622882Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.4623094Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4623253Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4623363Z unimplemented []
2025-12-04T09:45:15.4623489Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4624178Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.4624964Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4625143Z graph_break []
2025-12-04T09:45:15.4625274Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4625886Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.4626462Z   current_size = base.storage().size()
2025-12-04T09:45:15.4626584Z Autotune Choices Stats:
2025-12-04T09:45:15.4627395Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.4628329Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4628619Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4628937Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4629771Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4631352Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4632600Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4633889Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4635144Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4636405Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4637736Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4638982Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4640226Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4641513Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4642286Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.4642496Z Autotune Choices Stats:
2025-12-04T09:45:15.4643366Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.4644392Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4644817Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4645316Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4646267Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4647559Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4648844Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4650125Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4651486Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4653546Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4654845Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4656142Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4657428Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4658718Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4659503Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.4659748Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4659905Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4660016Z unimplemented []
2025-12-04T09:45:15.4660136Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4660334Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4661106Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4661761Z graph_break []
2025-12-04T09:45:15.4661893Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4662049Z Autotune Choices Stats:
2025-12-04T09:45:15.4662855Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.4663770Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4664049Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4664358Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4665170Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4666422Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4667668Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4668906Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4670161Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4671441Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4672689Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4673943Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4675180Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4676423Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4677192Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.4677397Z Autotune Choices Stats:
2025-12-04T09:45:15.4678214Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.4679260Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4679693Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4680175Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4681143Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4682435Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4683714Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4684991Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4686271Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4687583Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4688877Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4690156Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4691505Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4692794Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4693581Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.4693822Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4693987Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4694096Z unimplemented []
2025-12-04T09:45:15.4694219Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4694420Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4695137Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4695779Z graph_break []
2025-12-04T09:45:15.4695909Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4696064Z Autotune Choices Stats:
2025-12-04T09:45:15.4696901Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.4697813Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4698093Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4698408Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4699249Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4700540Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4701795Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4703039Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4704293Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4705563Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4706826Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4708062Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4709333Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4710655Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4711425Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.4711632Z Autotune Choices Stats:
2025-12-04T09:45:15.4712460Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.4713460Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4713879Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4714391Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4715356Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4716646Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4717941Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4719220Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4720563Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4721854Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4723173Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4724469Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4725759Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4727067Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4727860Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.4728103Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4728259Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4728370Z unimplemented []
2025-12-04T09:45:15.4728490Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4728685Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4729396Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.4730037Z graph_break []
2025-12-04T09:45:15.4730169Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4730323Z Autotune Choices Stats:
2025-12-04T09:45:15.4731160Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.4732064Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4732396Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4732707Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4733517Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4734781Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4736014Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4737256Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4738495Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4739739Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4741065Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4742331Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4743587Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4744849Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4745616Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.4745821Z Autotune Choices Stats:
2025-12-04T09:45:15.4746660Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.4747663Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4748086Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4748566Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4749534Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4750873Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4752157Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4753448Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4754750Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4756037Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4757312Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4758627Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4759924Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4761234Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4762020Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.4762261Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4762419Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4762529Z unimplemented []
2025-12-04T09:45:15.4762650Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4762848Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4763561Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4764202Z graph_break []
2025-12-04T09:45:15.4764334Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4764487Z Autotune Choices Stats:
2025-12-04T09:45:15.4765290Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.4766196Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4766478Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4766792Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4767638Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4768920Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4770197Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4771477Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4772718Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4773965Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4775216Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4776512Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4777754Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4779013Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4779785Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.4779988Z Autotune Choices Stats:
2025-12-04T09:45:15.4780850Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.4781858Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4782278Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4782761Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4783708Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4785019Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4786303Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4787592Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4788878Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4790172Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4791508Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4792785Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4794092Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4795387Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4796186Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.4796424Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4796579Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4796690Z unimplemented []
2025-12-04T09:45:15.4796809Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4797007Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4797724Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.4798362Z graph_break []
2025-12-04T09:45:15.4798498Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4798652Z Autotune Choices Stats:
2025-12-04T09:45:15.4799457Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.4800365Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4800691Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4801002Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4801807Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4803084Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4804332Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4805594Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4806840Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4808081Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4809326Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4810607Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4811901Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4813159Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4813944Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.4814147Z Autotune Choices Stats:
2025-12-04T09:45:15.4814975Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.4815990Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4816414Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4816889Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4817829Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4819112Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4820477Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4821799Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4823095Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4824399Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4825694Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4826980Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4828261Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4829585Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4830389Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.4830667Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4830824Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4830936Z unimplemented []
2025-12-04T09:45:15.4831056Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4831255Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4831996Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4832635Z graph_break []
2025-12-04T09:45:15.4832767Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4832923Z Autotune Choices Stats:
2025-12-04T09:45:15.4833732Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.4834632Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4834914Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4835223Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4836037Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4837290Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4838584Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4839829Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4841154Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4842400Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4843667Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4844929Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4846180Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4847458Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4848247Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.4848455Z Autotune Choices Stats:
2025-12-04T09:45:15.4849290Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.4850313Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4850782Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4851261Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4852207Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4853488Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4854772Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4856094Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4857390Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4858692Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4859986Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4861300Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4862589Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4863881Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4864695Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.4864948Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4865106Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4865216Z unimplemented []
2025-12-04T09:45:15.4865335Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4865533Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4866252Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.4866910Z graph_break []
2025-12-04T09:45:15.4867045Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4867202Z Autotune Choices Stats:
2025-12-04T09:45:15.4868009Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.4868906Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4869187Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4869499Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4870307Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4871599Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4872847Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4874162Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4875404Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4876661Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4877926Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4879189Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4880466Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4881711Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4882523Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.4882729Z Autotune Choices Stats:
2025-12-04T09:45:15.4883552Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.4884564Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4884999Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4885477Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4886421Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4887700Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4888987Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4890265Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4891596Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4892899Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4894201Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4895477Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4896784Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4898084Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4898873Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.4899115Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4899272Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4899383Z unimplemented []
2025-12-04T09:45:15.4899503Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4899705Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4900495Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.4901137Z graph_break []
2025-12-04T09:45:15.4901268Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4901426Z Autotune Choices Stats:
2025-12-04T09:45:15.4902241Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.4903157Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4903437Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4903746Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4904567Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4905847Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4907101Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4908372Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4909625Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4910905Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4912174Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4913437Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4914683Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4915931Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4916702Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.4916908Z Autotune Choices Stats:
2025-12-04T09:45:15.4917760Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.4918781Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4919205Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4919684Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4920677Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4921965Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4923243Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4924524Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4925814Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4927147Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4928441Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4929738Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4931056Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4932343Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4933134Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.4933396Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.4933577Z Traceback (most recent call last):
2025-12-04T09:45:15.4933814Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.4934039Z     self.assertTrue(
2025-12-04T09:45:15.4934208Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.4934396Z     raise self.failureException(msg)
2025-12-04T09:45:15.4934608Z AssertionError: False is not true : Log file /tmp/tmp9gltujgb/flex_attention_configs.json was not created
2025-12-04T09:45:15.4934769Z 
2025-12-04T09:45:15.4934849Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.4935124Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.4935345Z 
2025-12-04T09:45:15.4935475Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.4935677Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4935832Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4935942Z unimplemented []
2025-12-04T09:45:15.4936064Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4936740Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.4937467Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4937642Z graph_break []
2025-12-04T09:45:15.4937774Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4938381Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.4938950Z   current_size = base.storage().size()
2025-12-04T09:45:15.4939072Z Autotune Choices Stats:
2025-12-04T09:45:15.4939890Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.4940832Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4941110Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4941419Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4942238Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4943480Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4944764Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4946006Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4947254Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4948498Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4949749Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4951034Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4952294Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4953564Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4954341Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.4954547Z Autotune Choices Stats:
2025-12-04T09:45:15.4955370Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.4956396Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4961832Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4962331Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4963282Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4964558Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4965846Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4967194Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4968492Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4969792Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4971122Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4972405Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4973680Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4974959Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4975786Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.4976042Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.4976200Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.4976310Z unimplemented []
2025-12-04T09:45:15.4976430Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.4976627Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.4977344Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.4978010Z graph_break []
2025-12-04T09:45:15.4978142Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.4978296Z Autotune Choices Stats:
2025-12-04T09:45:15.4979108Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.4980006Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4980286Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4980638Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4981445Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4982694Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4983933Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4985214Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4986450Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4987698Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.4988941Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4990185Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4991479Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4992723Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4993533Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.4993738Z Autotune Choices Stats:
2025-12-04T09:45:15.4994555Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.4995564Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.4996002Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.4996477Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.4997421Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4998698Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.4999969Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5001272Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5002576Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5003879Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5005178Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5006473Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5007757Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5009201Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5009989Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.5010225Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5010380Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5010516Z unimplemented []
2025-12-04T09:45:15.5010631Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5010825Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5011600Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5012235Z graph_break []
2025-12-04T09:45:15.5012364Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5012517Z Autotune Choices Stats:
2025-12-04T09:45:15.5013317Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.5014233Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5014510Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5014817Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5015627Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5016867Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5018097Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5019356Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5020632Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5021882Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5023154Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5024392Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5025643Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5026904Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5027675Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.5027878Z Autotune Choices Stats:
2025-12-04T09:45:15.5028722Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.5029737Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5030156Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5030667Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5031621Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5032909Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5034194Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5035488Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5036771Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5038105Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5039398Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5040722Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5041996Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5043286Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5044074Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.5044151Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5044195Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5044232Z unimplemented []
2025-12-04T09:45:15.5044292Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5044392Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5044962Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5045014Z graph_break []
2025-12-04T09:45:15.5045118Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5045158Z Autotune Choices Stats:
2025-12-04T09:45:15.5045897Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.5046025Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5046156Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5046321Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5046934Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5047532Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5048139Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5048752Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5049393Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5050016Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5050653Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5051270Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5051872Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5052476Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5052609Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.5052651Z Autotune Choices Stats:
2025-12-04T09:45:15.5053409Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.5053627Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5053829Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5054106Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5054734Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5055369Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5055993Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5056617Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5057240Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5057907Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5058542Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5059275Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5059915Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5060592Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5060724Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.5060799Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5060841Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5060880Z unimplemented []
2025-12-04T09:45:15.5060941Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5061043Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5061620Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5061659Z graph_break []
2025-12-04T09:45:15.5061731Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5061771Z Autotune Choices Stats:
2025-12-04T09:45:15.5062553Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.5062692Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5062809Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5062968Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5063582Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5064209Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5064816Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5065436Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5066046Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5066668Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5067280Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5067901Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5068515Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5069117Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5069245Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.5069286Z Autotune Choices Stats:
2025-12-04T09:45:15.5070053Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.5070271Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5070474Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5070779Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5071423Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5072073Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5072706Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5073325Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5073953Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5074581Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5075241Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5075875Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5076506Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5077150Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5077280Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.5077356Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5077399Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5077436Z unimplemented []
2025-12-04T09:45:15.5077496Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5077595Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5078173Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5078212Z graph_break []
2025-12-04T09:45:15.5078286Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5078325Z Autotune Choices Stats:
2025-12-04T09:45:15.5079069Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.5079197Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5079341Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5079502Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5080116Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5080765Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5081394Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5082017Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5082635Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5083236Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5083866Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5084482Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5085087Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5085706Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5085835Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.5085878Z Autotune Choices Stats:
2025-12-04T09:45:15.5086632Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.5086849Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5087018Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5087296Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5087943Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5088567Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5089185Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5089821Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5090488Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5091113Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5091745Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5092404Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5093043Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5093671Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5093815Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.5093890Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5093931Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5093970Z unimplemented []
2025-12-04T09:45:15.5094031Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5094130Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5094707Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5094743Z graph_break []
2025-12-04T09:45:15.5094815Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5094854Z Autotune Choices Stats:
2025-12-04T09:45:15.5095598Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.5095726Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5095839Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5096001Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5096639Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5097250Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5097864Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5098470Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5099077Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5099688Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5100288Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5100950Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5101564Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5102165Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5102307Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.5102345Z Autotune Choices Stats:
2025-12-04T09:45:15.5103103Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.5103320Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5103484Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5103759Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5104395Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5105037Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5105668Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5106306Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5106947Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5107577Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5108197Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5108822Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5109476Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5110102Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5110231Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.5110317Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5110357Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5110394Z unimplemented []
2025-12-04T09:45:15.5110491Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5110591Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5111168Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5111206Z graph_break []
2025-12-04T09:45:15.5111281Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5111321Z Autotune Choices Stats:
2025-12-04T09:45:15.5112064Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.5112190Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5112307Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5112464Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5113074Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5113715Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5114330Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5114948Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5115554Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5116162Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5116768Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5117373Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5118000Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5118615Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5118753Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.5118798Z Autotune Choices Stats:
2025-12-04T09:45:15.5119564Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.5119781Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5119955Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5120230Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5120904Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5121525Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5122168Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5122799Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5123427Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5124057Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5124683Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5125308Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5125935Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5126604Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5126743Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.5126815Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5126857Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5126894Z unimplemented []
2025-12-04T09:45:15.5126955Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5127054Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5127644Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5127680Z graph_break []
2025-12-04T09:45:15.5127754Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5127793Z Autotune Choices Stats:
2025-12-04T09:45:15.5128533Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.5128662Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5128775Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5128935Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5129546Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5130149Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5130804Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5131411Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5132023Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5132625Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5133233Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5133855Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5134460Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5135080Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5135220Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.5135258Z Autotune Choices Stats:
2025-12-04T09:45:15.5136017Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.5136247Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5136411Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5136685Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5137323Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5137949Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5138564Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5139197Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5139835Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5140516Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5141137Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5141766Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5142393Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5143016Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5143143Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.5143261Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5143302Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5143340Z unimplemented []
2025-12-04T09:45:15.5143399Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5143499Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5144075Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5144126Z graph_break []
2025-12-04T09:45:15.5144201Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5144241Z Autotune Choices Stats:
2025-12-04T09:45:15.5144982Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.5145110Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5145227Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5145388Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5146011Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5146621Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5147232Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5147861Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5148465Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5149077Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5149675Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5150280Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5150934Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5151535Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5151663Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.5151743Z Autotune Choices Stats:
2025-12-04T09:45:15.5152500Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.5152720Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5152903Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5153181Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5153813Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5154437Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5155071Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5155685Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5156330Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5156972Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5157606Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5158236Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5158862Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5159510Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5159643Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.5159734Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.5159780Z Traceback (most recent call last):
2025-12-04T09:45:15.5159935Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.5159975Z     self.assertTrue(
2025-12-04T09:45:15.5160079Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.5160139Z     raise self.failureException(msg)
2025-12-04T09:45:15.5160288Z AssertionError: False is not true : Log file /tmp/tmp3ho9z2ol/flex_attention_configs.json was not created
2025-12-04T09:45:15.5160291Z 
2025-12-04T09:45:15.5160367Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.5160556Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.5160558Z 
2025-12-04T09:45:15.5160647Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.5160720Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5160764Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5160801Z unimplemented []
2025-12-04T09:45:15.5160862Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5161438Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.5161553Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5161590Z graph_break []
2025-12-04T09:45:15.5161661Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5162153Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.5162203Z   current_size = base.storage().size()
2025-12-04T09:45:15.5162244Z Autotune Choices Stats:
2025-12-04T09:45:15.5162982Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.5163110Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5163228Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5163388Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5164007Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5164638Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5165249Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5165856Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5166456Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5167059Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5167661Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5168260Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5168876Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5169490Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5169628Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.5169670Z Autotune Choices Stats:
2025-12-04T09:45:15.5170464Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.5170683Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5170851Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5171128Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5171761Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5172391Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5173039Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5173671Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5174313Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5174954Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5175574Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5176203Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5176847Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5177495Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5177633Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.5177706Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5177746Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5177783Z unimplemented []
2025-12-04T09:45:15.5177842Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5177941Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5178531Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5178567Z graph_break []
2025-12-04T09:45:15.5178640Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5178678Z Autotune Choices Stats:
2025-12-04T09:45:15.5179420Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.5179549Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5179663Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5179822Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5180462Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5181071Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5181696Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5182308Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5182923Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5183530Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5184136Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5184740Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5185348Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5185972Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5186114Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.5186154Z Autotune Choices Stats:
2025-12-04T09:45:15.5186917Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.5187146Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5187312Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5187589Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5188220Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5188843Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5189480Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5190146Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5190816Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5191460Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5192082Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5192708Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5193440Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5194059Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5194187Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.5194303Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5194345Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5194382Z unimplemented []
2025-12-04T09:45:15.5194442Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5194540Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5195112Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5195162Z graph_break []
2025-12-04T09:45:15.5195233Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5195275Z Autotune Choices Stats:
2025-12-04T09:45:15.5196011Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.5196138Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5196255Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5196415Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5197029Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5197650Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5198257Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5198890Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5199496Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5200110Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5200756Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5201355Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5201960Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5202561Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5202690Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.5202751Z Autotune Choices Stats:
2025-12-04T09:45:15.5203536Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.5203753Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5203942Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5204222Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5204850Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5205471Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5206097Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5206725Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5207380Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5208016Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5208650Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5209265Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5209908Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5210585Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5210714Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.5210787Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5210828Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5210866Z unimplemented []
2025-12-04T09:45:15.5210926Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5211025Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5211634Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5211685Z graph_break []
2025-12-04T09:45:15.5211759Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5211799Z Autotune Choices Stats:
2025-12-04T09:45:15.5212540Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.5212682Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5212797Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5212957Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5213562Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5214166Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5214768Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5215363Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5216084Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5216691Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5217310Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5217914Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5218528Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5219151Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5219282Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.5219322Z Autotune Choices Stats:
2025-12-04T09:45:15.5220103Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.5220338Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5220542Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5220824Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5221464Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5222081Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5222703Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5223325Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5223952Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5224595Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5225230Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5225867Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5226511Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5227135Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5227265Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.5227340Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5227385Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5227422Z unimplemented []
2025-12-04T09:45:15.5227483Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5227582Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5228152Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5228190Z graph_break []
2025-12-04T09:45:15.5228261Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5228342Z Autotune Choices Stats:
2025-12-04T09:45:15.5229081Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.5229209Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5229335Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5229497Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5230111Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5230747Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5231357Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5231966Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5232585Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5233225Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5233836Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5234450Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5235069Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5235674Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5235805Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.5235849Z Autotune Choices Stats:
2025-12-04T09:45:15.5236599Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.5236819Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5237015Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5237301Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5237938Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5238573Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5239213Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5239839Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5240502Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5241129Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5241789Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5242411Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5243053Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5243683Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5243812Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.5243887Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5243928Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5243965Z unimplemented []
2025-12-04T09:45:15.5244024Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5244124Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5244702Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5244741Z graph_break []
2025-12-04T09:45:15.5244814Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5244853Z Autotune Choices Stats:
2025-12-04T09:45:15.5245620Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.5245755Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5245874Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5246035Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5246646Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5247261Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5247869Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5248474Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5249077Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5249701Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5250315Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5250953Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5251566Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5252171Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5252299Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.5252340Z Autotune Choices Stats:
2025-12-04T09:45:15.5253112Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.5253330Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5253496Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5253771Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5254447Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5255073Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5255705Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5256331Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5256961Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5257605Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5258227Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5258882Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5259505Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5260136Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5260264Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.5260338Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5260383Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5260425Z unimplemented []
2025-12-04T09:45:15.5260513Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5260612Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5261182Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5261218Z graph_break []
2025-12-04T09:45:15.5261293Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5261334Z Autotune Choices Stats:
2025-12-04T09:45:15.5262082Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.5262212Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5262353Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5262527Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5263140Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5263745Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5264356Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5264962Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5265571Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5266173Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5266809Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5267426Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5268037Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5268653Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5268782Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.5268823Z Autotune Choices Stats:
2025-12-04T09:45:15.5269576Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.5269795Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5269962Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5270245Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5270981Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5271641Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5272268Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5272907Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5273552Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5274179Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5274809Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5275483Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5276118Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5276758Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5276899Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.5276974Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5277015Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5277054Z unimplemented []
2025-12-04T09:45:15.5277113Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5277213Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5277789Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5277831Z graph_break []
2025-12-04T09:45:15.5277904Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5277947Z Autotune Choices Stats:
2025-12-04T09:45:15.5278692Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.5278819Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5278937Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5279097Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5279730Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5280343Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5280986Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5281605Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5282210Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5282815Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5283441Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5284071Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5284691Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5285292Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5285432Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.5285475Z Autotune Choices Stats:
2025-12-04T09:45:15.5286240Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.5286458Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5286625Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5286900Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5287539Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5288200Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5288832Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5289452Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5290095Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5290764Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5291391Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5292017Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5292674Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5293309Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5293440Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.5293526Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5293571Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5293607Z unimplemented []
2025-12-04T09:45:15.5293669Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5293768Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5294343Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5294379Z graph_break []
2025-12-04T09:45:15.5294453Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5294494Z Autotune Choices Stats:
2025-12-04T09:45:15.5295231Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.5295358Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5295472Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5295635Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5296251Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5296877Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5297488Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5298104Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5298723Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5299330Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5299959Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5300593Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5301237Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5301853Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5301984Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.5302036Z Autotune Choices Stats:
2025-12-04T09:45:15.5302795Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.5303013Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5303182Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5303458Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5304090Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5304708Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5305351Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5305990Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5306622Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5307261Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5307885Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5308515Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5309137Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5309779Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5309919Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.5309992Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5310034Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5310071Z unimplemented []
2025-12-04T09:45:15.5310130Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5310230Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5310834Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5310888Z graph_break []
2025-12-04T09:45:15.5310959Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5311000Z Autotune Choices Stats:
2025-12-04T09:45:15.5311746Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.5311874Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5311989Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5312150Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5312783Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5313389Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5314022Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5314637Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5315255Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5315862Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5316484Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5317086Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5317697Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5318331Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5318469Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.5318512Z Autotune Choices Stats:
2025-12-04T09:45:15.5319271Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.5319500Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5319667Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5319942Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5320618Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5321239Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5321871Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5322527Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5323179Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5323812Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5324438Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5325071Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5325771Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5326393Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5326522Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.5326595Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5326671Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5326708Z unimplemented []
2025-12-04T09:45:15.5326768Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5326866Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5327441Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5327477Z graph_break []
2025-12-04T09:45:15.5327562Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5327603Z Autotune Choices Stats:
2025-12-04T09:45:15.5328345Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.5328472Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5328586Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5328751Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5329380Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5329983Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5330637Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5331268Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5331887Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5332509Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5333115Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5333737Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5334358Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5334976Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5335104Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.5335144Z Autotune Choices Stats:
2025-12-04T09:45:15.5335937Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.5336156Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5336333Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5336613Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5337246Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5337934Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5338557Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5339199Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5339855Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5340523Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5341155Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5341785Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5342405Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5343031Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5343161Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.5343254Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.5343301Z Traceback (most recent call last):
2025-12-04T09:45:15.5343453Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.5343492Z     self.assertTrue(
2025-12-04T09:45:15.5343597Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.5343646Z     raise self.failureException(msg)
2025-12-04T09:45:15.5343800Z AssertionError: False is not true : Log file /tmp/tmp081kh257/flex_attention_configs.json was not created
2025-12-04T09:45:15.5343816Z 
2025-12-04T09:45:15.5343893Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.5344060Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.5344063Z 
2025-12-04T09:45:15.5344153Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.5344228Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5344270Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5344308Z unimplemented []
2025-12-04T09:45:15.5344367Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5344948Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.5345059Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5345094Z graph_break []
2025-12-04T09:45:15.5345168Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5345665Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.5345717Z   current_size = base.storage().size()
2025-12-04T09:45:15.5345758Z Autotune Choices Stats:
2025-12-04T09:45:15.5346498Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.5346626Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5346743Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5346906Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5347524Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5348148Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5348769Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5349385Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5349986Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5350620Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5351229Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5351834Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5352465Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5353078Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5353210Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.5353263Z Autotune Choices Stats:
2025-12-04T09:45:15.5354024Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.5354245Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5354414Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5354692Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5355323Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5355953Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5356596Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5357228Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5357855Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5358492Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5359113Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5359740Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5360368Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5361035Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5361179Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.5361252Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5361294Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5361333Z unimplemented []
2025-12-04T09:45:15.5361395Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5361494Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5362070Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5362129Z graph_break []
2025-12-04T09:45:15.5362202Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5362243Z Autotune Choices Stats:
2025-12-04T09:45:15.5362981Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.5363111Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5363225Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5363384Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5363999Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5364603Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5365239Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5365847Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5366466Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5367082Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5367691Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5368303Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5368911Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5369528Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5369669Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.5369710Z Autotune Choices Stats:
2025-12-04T09:45:15.5370499Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.5370742Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5370916Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5371192Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5371830Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5372456Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5373079Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5373733Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5374382Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5375025Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5375659Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5376288Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5376922Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5377550Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5377680Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.5377756Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5377809Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5377873Z unimplemented []
2025-12-04T09:45:15.5377934Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5378036Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5378603Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5378642Z graph_break []
2025-12-04T09:45:15.5378716Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5378768Z Autotune Choices Stats:
2025-12-04T09:45:15.5379517Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.5379647Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5379765Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5379998Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5380642Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5381242Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5381854Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5382520Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5383134Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5383758Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5384365Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5384978Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5385589Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5386198Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5386332Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.5386373Z Autotune Choices Stats:
2025-12-04T09:45:15.5387161Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.5387380Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5387549Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5387841Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5388483Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5389107Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5389727Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5390352Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5391045Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5391686Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5392332Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5392962Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5393585Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5394210Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5394341Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.5394414Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5394460Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5394499Z unimplemented []
2025-12-04T09:45:15.5394562Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5394661Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5395265Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5397041Z graph_break []
2025-12-04T09:45:15.5397113Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5397155Z Autotune Choices Stats:
2025-12-04T09:45:15.5397900Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.5398049Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5398163Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5398325Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5398946Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5399554Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5400181Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5400841Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5401487Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5402107Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5402728Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5403338Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5403946Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5404550Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5404681Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.5404723Z Autotune Choices Stats:
2025-12-04T09:45:15.5405503Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.5405734Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5405899Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5406180Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5406821Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5407448Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5408072Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5408704Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5409337Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5409993Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5410661Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5411306Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5411937Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5412569Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5412701Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.5412777Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5412820Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5412859Z unimplemented []
2025-12-04T09:45:15.5412919Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5413020Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5413604Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5413644Z graph_break []
2025-12-04T09:45:15.5413717Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5413773Z Autotune Choices Stats:
2025-12-04T09:45:15.5414551Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.5414680Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5414797Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5414971Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5415598Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5416214Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5416820Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5417426Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5418037Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5418670Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5419276Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5419892Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5420547Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5421154Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5421284Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.5421328Z Autotune Choices Stats:
2025-12-04T09:45:15.5422093Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.5422313Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5422504Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5422794Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5423427Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5424065Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5424691Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5425315Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5425948Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5426577Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5427227Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5427858Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5428496Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5429120Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5429251Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.5429323Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5429368Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5429405Z unimplemented []
2025-12-04T09:45:15.5429467Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5429568Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5430145Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5430183Z graph_break []
2025-12-04T09:45:15.5430258Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5430297Z Autotune Choices Stats:
2025-12-04T09:45:15.5431109Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.5431249Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5431363Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5431523Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5432143Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5432758Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5433361Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5433963Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5434581Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5435184Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5435821Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5436425Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5437042Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5437660Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5437790Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.5437830Z Autotune Choices Stats:
2025-12-04T09:45:15.5438586Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.5438808Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5438974Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5439257Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5439921Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5440589Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5441222Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5441843Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5442475Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5443098Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5443718Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5444396Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5445020Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5445659Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5445786Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.5445860Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5445903Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5445942Z unimplemented []
2025-12-04T09:45:15.5446001Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5446102Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5446676Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5446714Z graph_break []
2025-12-04T09:45:15.5446786Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5446829Z Autotune Choices Stats:
2025-12-04T09:45:15.5447578Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.5447705Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5447819Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5448016Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5448636Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5449245Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5449860Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5450519Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5451116Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5451727Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5452388Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5453005Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5453609Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5454221Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5454352Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.5454393Z Autotune Choices Stats:
2025-12-04T09:45:15.5455164Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.5455382Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5455551Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5455827Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5456457Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5457113Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5457736Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5458359Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5458989Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5459629Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5460251Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5460933Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5461611Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5462235Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5462385Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.5462458Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5462501Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5462538Z unimplemented []
2025-12-04T09:45:15.5462599Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5462698Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5463282Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5463319Z graph_break []
2025-12-04T09:45:15.5463395Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5463433Z Autotune Choices Stats:
2025-12-04T09:45:15.5464179Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.5464306Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5464420Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5464581Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5465218Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5465835Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5466442Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5467060Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5467667Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5468273Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5468899Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5469527Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5470142Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5470858Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5471007Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.5471049Z Autotune Choices Stats:
2025-12-04T09:45:15.5471809Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.5477026Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5477235Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5477543Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5478270Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5478922Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5479631Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5480264Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5480986Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5481634Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5482400Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5483034Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5483698Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5484354Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5484489Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.5484574Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5484638Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5484680Z unimplemented []
2025-12-04T09:45:15.5484747Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5484853Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5485456Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5485505Z graph_break []
2025-12-04T09:45:15.5485585Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5485630Z Autotune Choices Stats:
2025-12-04T09:45:15.5486409Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.5486541Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5486661Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5486833Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5487452Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5488082Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5488747Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5489376Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5489993Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5490631Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5491241Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5491870Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5492518Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5493140Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5493277Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.5493321Z Autotune Choices Stats:
2025-12-04T09:45:15.5494097Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.5494322Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5494496Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5494781Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5495416Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5496046Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5496690Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5497334Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5497986Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5498631Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5499264Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5499904Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5500570Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5501244Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5501392Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.5501470Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5501520Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5501558Z unimplemented []
2025-12-04T09:45:15.5501623Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5501726Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5502305Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5502360Z graph_break []
2025-12-04T09:45:15.5502433Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5502478Z Autotune Choices Stats:
2025-12-04T09:45:15.5505561Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.5505771Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5505890Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5506058Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5506700Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5507315Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5507961Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5508634Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5509281Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5509942Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5510660Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5511271Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5511882Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5512518Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5512665Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.5512714Z Autotune Choices Stats:
2025-12-04T09:45:15.5513479Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.5513719Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5513892Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5514173Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5535418Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5536057Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5536689Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5537356Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5538001Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5538654Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5539291Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5539926Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5540612Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5541242Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5541373Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.5541457Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5541506Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5541576Z unimplemented []
2025-12-04T09:45:15.5541668Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5541778Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5542355Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5542396Z graph_break []
2025-12-04T09:45:15.5542478Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5542533Z Autotune Choices Stats:
2025-12-04T09:45:15.5543292Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.5543423Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5543546Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5543719Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5544333Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5544948Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5545562Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5546206Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5546831Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5547449Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5548077Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5548685Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5549314Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5549926Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5550063Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.5550107Z Autotune Choices Stats:
2025-12-04T09:45:15.5550945Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.5551180Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5551356Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5551659Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5552290Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5552922Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5553560Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5554191Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5554835Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5555477Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5556104Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5556748Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5557374Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5558020Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5558151Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.5558223Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5558267Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5558304Z unimplemented []
2025-12-04T09:45:15.5558364Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5558465Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5559064Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5559110Z graph_break []
2025-12-04T09:45:15.5559183Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5559222Z Autotune Choices Stats:
2025-12-04T09:45:15.5559967Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.5560105Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5560220Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5560380Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5561015Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5561612Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5562218Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5562820Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5563445Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5564059Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5564682Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5565280Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5565876Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5566482Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5566613Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.5566652Z Autotune Choices Stats:
2025-12-04T09:45:15.5567432Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.5567660Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5567825Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5568104Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5568737Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5569369Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5569991Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5570666Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5571297Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5571962Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5572599Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5573240Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5573869Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5574491Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5574619Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.5574711Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.5574759Z Traceback (most recent call last):
2025-12-04T09:45:15.5574919Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.5574958Z     self.assertTrue(
2025-12-04T09:45:15.5575067Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.5575115Z     raise self.failureException(msg)
2025-12-04T09:45:15.5575244Z AssertionError: False is not true : Log file /tmp/tmpxkkpz8q7/flex_attention_configs.json was not created
2025-12-04T09:45:15.5575247Z 
2025-12-04T09:45:15.5575323Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.5575488Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.5575493Z 
2025-12-04T09:45:15.5575580Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.5575657Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5575729Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5575767Z unimplemented []
2025-12-04T09:45:15.5575827Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5576407Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.5576506Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5576541Z graph_break []
2025-12-04T09:45:15.5576615Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5577122Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.5577170Z   current_size = base.storage().size()
2025-12-04T09:45:15.5577210Z Autotune Choices Stats:
2025-12-04T09:45:15.5577958Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.5578089Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5578202Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5578362Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5578968Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5579567Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5580182Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5580827Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5581440Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5582047Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5582673Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5583282Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5583891Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5584513Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5584655Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.5584693Z Autotune Choices Stats:
2025-12-04T09:45:15.5585453Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.5585683Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5585849Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5586128Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5586761Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5587383Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5588012Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5588661Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5589295Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5589923Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5590594Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5591226Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5591846Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5592460Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5592589Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.5592662Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5592706Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5592759Z unimplemented []
2025-12-04T09:45:15.5592843Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5592943Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5593511Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5593548Z graph_break []
2025-12-04T09:45:15.5593620Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5593671Z Autotune Choices Stats:
2025-12-04T09:45:15.5594411Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.5594538Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5594652Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5594817Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5595429Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5596046Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5596639Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5597258Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5597867Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5598478Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5599086Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5599689Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5600284Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5600919Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5601046Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.5601085Z Autotune Choices Stats:
2025-12-04T09:45:15.5601875Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.5602107Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5602274Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5602563Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5603198Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5603823Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5604438Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5605060Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5605719Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5606352Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5606987Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5607622Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5608252Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5608873Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5609003Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.5609078Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5609120Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5609158Z unimplemented []
2025-12-04T09:45:15.5609217Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5609317Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5609906Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5609952Z graph_break []
2025-12-04T09:45:15.5610026Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5610064Z Autotune Choices Stats:
2025-12-04T09:45:15.5610838Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.5610981Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5611095Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5611253Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5611873Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5612477Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5613081Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5613679Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5614308Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5614935Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5615547Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5616158Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5616764Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5617363Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5617493Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.5617532Z Autotune Choices Stats:
2025-12-04T09:45:15.5618311Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.5618539Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5618707Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5618983Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5619612Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5620251Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5620907Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5621532Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5622174Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5622827Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5623469Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5624109Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5624733Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5625366Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5625495Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.5625569Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5625612Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5625649Z unimplemented []
2025-12-04T09:45:15.5625710Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5625808Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5626378Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5626415Z graph_break []
2025-12-04T09:45:15.5626486Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5626526Z Autotune Choices Stats:
2025-12-04T09:45:15.5627279Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.5627419Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5627534Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5627706Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5628322Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5628923Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5629524Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5630130Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5630767Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5631385Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5632010Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5632625Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5633231Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5633826Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5633954Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.5633994Z Autotune Choices Stats:
2025-12-04T09:45:15.5634761Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.5634980Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5635165Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5635453Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5636082Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5636719Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5637346Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5637972Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5638606Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5639240Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5639875Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5640558Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5641201Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5641824Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5641955Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.5642029Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5642069Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5642105Z unimplemented []
2025-12-04T09:45:15.5642163Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5642266Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5642844Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5642883Z graph_break []
2025-12-04T09:45:15.5642956Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5642996Z Autotune Choices Stats:
2025-12-04T09:45:15.5643758Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.5643897Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5644011Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5644172Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5644786Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5645405Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5646030Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5646649Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5647253Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5647859Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5648516Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5649128Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5649743Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5650345Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5650515Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.5650553Z Autotune Choices Stats:
2025-12-04T09:45:15.5651315Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.5651536Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5651701Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5651978Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5652640Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5653274Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5653914Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5654535Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5655159Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5655800Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5656430Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5657093Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5657727Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5658361Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5658488Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.5658560Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5658600Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5658637Z unimplemented []
2025-12-04T09:45:15.5658699Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5658798Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5659364Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5659401Z graph_break []
2025-12-04T09:45:15.5659473Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5659512Z Autotune Choices Stats:
2025-12-04T09:45:15.5660255Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.5660381Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5660531Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5660717Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5661340Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5661959Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5662581Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5663186Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5663788Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5664407Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5665015Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5665649Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5666251Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5666865Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5666992Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.5667031Z Autotune Choices Stats:
2025-12-04T09:45:15.5667794Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.5668014Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5668181Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5668462Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5669090Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5669758Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5670385Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5671075Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5671706Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5672328Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5672950Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5673581Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5674244Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5674869Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5675010Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.5675082Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5675123Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5675159Z unimplemented []
2025-12-04T09:45:15.5675220Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5675320Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5675897Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5675934Z graph_break []
2025-12-04T09:45:15.5676006Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5676044Z Autotune Choices Stats:
2025-12-04T09:45:15.5676787Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.5676915Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5677028Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5677188Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5677825Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5678435Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5679054Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5679668Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5680274Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5680924Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5681531Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5682141Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5682789Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5683391Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5683538Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.5683576Z Autotune Choices Stats:
2025-12-04T09:45:15.5684347Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.5684568Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5684732Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5685011Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5685645Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5686266Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5686915Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5687537Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5688173Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5688801Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5689416Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5690050Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5690721Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5691381Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5691509Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.5691582Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5691622Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5691672Z unimplemented []
2025-12-04T09:45:15.5691732Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5691833Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5692407Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5692444Z graph_break []
2025-12-04T09:45:15.5692515Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5692554Z Autotune Choices Stats:
2025-12-04T09:45:15.5693298Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.5693425Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5693539Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5693698Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5694304Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5694929Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5695546Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5696149Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5696758Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5697355Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5697956Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5698563Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5699185Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5699806Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5699933Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.5699973Z Autotune Choices Stats:
2025-12-04T09:45:15.5700784Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.5701018Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5701185Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5701462Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5702093Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5702718Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5703356Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5704009Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5704633Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5705268Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5705889Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5706521Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5707148Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5707789Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5707925Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.5707998Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5708039Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5708077Z unimplemented []
2025-12-04T09:45:15.5708135Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5708238Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5708811Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5708859Z graph_break []
2025-12-04T09:45:15.5708932Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5708970Z Autotune Choices Stats:
2025-12-04T09:45:15.5709710Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.5709837Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5709950Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5710110Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5710787Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5711391Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5712015Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5712625Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5713230Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5713840Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5714464Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5715072Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5715678Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5716300Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5716439Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.5716477Z Autotune Choices Stats:
2025-12-04T09:45:15.5717248Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.5717479Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5717643Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5717919Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5718553Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5719182Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5719799Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5720482Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5721120Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5721749Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5722381Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5723002Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5723628Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5724249Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5724375Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.5724447Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5724488Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5724524Z unimplemented []
2025-12-04T09:45:15.5724596Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5724714Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5725289Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5725326Z graph_break []
2025-12-04T09:45:15.5725396Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5725437Z Autotune Choices Stats:
2025-12-04T09:45:15.5726194Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.5726321Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5726435Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5726596Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5727214Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5727821Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5728420Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5729040Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5729653Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5730259Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5730905Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5731510Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5732116Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5732720Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5732848Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.5732888Z Autotune Choices Stats:
2025-12-04T09:45:15.5733683Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.5733912Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5734078Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5734373Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5735025Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5735651Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5736268Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5736889Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5737534Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5738167Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5738789Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5739425Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5740055Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5740718Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5740847Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.5740921Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5740961Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5740998Z unimplemented []
2025-12-04T09:45:15.5741055Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5741158Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5741766Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5741814Z graph_break []
2025-12-04T09:45:15.5741888Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5741926Z Autotune Choices Stats:
2025-12-04T09:45:15.5742671Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.5742811Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5742925Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5743085Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5743699Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5744305Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5744904Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5745504Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5746119Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5746727Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5747346Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5747952Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5748574Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5749170Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5749300Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.5749338Z Autotune Choices Stats:
2025-12-04T09:45:15.5750104Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.5750357Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5750568Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5750847Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5751483Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5752138Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5752758Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5753398Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5754028Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5754693Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5755325Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5755957Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5756595Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5757224Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5757352Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.5757424Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5757465Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5757502Z unimplemented []
2025-12-04T09:45:15.5757562Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5757662Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5758243Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5758278Z graph_break []
2025-12-04T09:45:15.5758350Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5758390Z Autotune Choices Stats:
2025-12-04T09:45:15.5759142Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.5759279Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5759393Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5759554Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5760179Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5760819Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5761424Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5762042Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5762658Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5763285Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5763906Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5764526Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5765136Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5765755Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5765883Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.5765922Z Autotune Choices Stats:
2025-12-04T09:45:15.5766687Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.5766906Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5767071Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5767379Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5768007Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5768630Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5769258Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5769880Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5770540Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5771175Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5771846Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5772484Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5773126Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5773757Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5773887Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.5773960Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5774000Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5774039Z unimplemented []
2025-12-04T09:45:15.5774098Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5774197Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5774774Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5774815Z graph_break []
2025-12-04T09:45:15.5774887Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5774925Z Autotune Choices Stats:
2025-12-04T09:45:15.5775669Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.5775825Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5775941Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5776103Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5776715Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5777331Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5777937Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5778543Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5779147Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5779747Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5780381Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5781031Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5781653Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5782254Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5782385Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:15.5782422Z Autotune Choices Stats:
2025-12-04T09:45:15.5783189Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.5783408Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5783573Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5783848Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5784501Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5785139Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5785771Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5786407Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5787052Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5787676Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5788312Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5788962Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5789594Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5790227Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5790356Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:15.5790493Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.5790540Z Traceback (most recent call last):
2025-12-04T09:45:15.5790692Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.5790735Z     self.assertTrue(
2025-12-04T09:45:15.5790837Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.5790886Z     raise self.failureException(msg)
2025-12-04T09:45:15.5791013Z AssertionError: False is not true : Log file /tmp/tmp9dxvkrhw/flex_attention_configs.json was not created
2025-12-04T09:45:15.5791016Z 
2025-12-04T09:45:15.5791092Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.5791257Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.5791261Z 
2025-12-04T09:45:15.5791351Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.5791427Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5791469Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5791507Z unimplemented []
2025-12-04T09:45:15.5791568Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5792145Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.5792243Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5792280Z graph_break []
2025-12-04T09:45:15.5792351Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5792871Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.5792932Z   current_size = base.storage().size()
2025-12-04T09:45:15.5792971Z Autotune Choices Stats:
2025-12-04T09:45:15.5793719Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.5793859Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5793974Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5794133Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5794745Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5795355Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5795959Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5796559Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5797186Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5797796Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5798412Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5799012Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5799609Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5800221Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5800352Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.5800392Z Autotune Choices Stats:
2025-12-04T09:45:15.5801180Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.5801434Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5801601Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5801875Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5802503Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5803141Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5803762Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5804387Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5805008Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5805652Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5806281Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5806926Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5807559Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5808185Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5808313Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.5808387Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5808429Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5808466Z unimplemented []
2025-12-04T09:45:15.5808526Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5808626Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5809206Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5809244Z graph_break []
2025-12-04T09:45:15.5809317Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5809356Z Autotune Choices Stats:
2025-12-04T09:45:15.5810117Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.5810255Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5810368Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5810568Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5811206Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5811808Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5812427Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5813026Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5813630Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5814254Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5814868Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5815482Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5816085Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5816690Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5816818Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.5816856Z Autotune Choices Stats:
2025-12-04T09:45:15.5817614Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.5817832Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5818000Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5818309Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5818939Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5819638Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5820272Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5820940Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5821565Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5822191Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5822843Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5823480Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5824122Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5824747Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5824879Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.5824952Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5824992Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5825029Z unimplemented []
2025-12-04T09:45:15.5825087Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5825188Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5825756Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5825795Z graph_break []
2025-12-04T09:45:15.5825866Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5825905Z Autotune Choices Stats:
2025-12-04T09:45:15.5826645Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.5826805Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5826920Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5827079Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5827683Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5828298Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5828904Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5829504Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5830104Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5830748Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5831376Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5831993Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5832609Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5833219Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5833348Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.5833390Z Autotune Choices Stats:
2025-12-04T09:45:15.5834147Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.5834366Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5834531Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5834805Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5835460Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5836098Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5836726Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5837354Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5838002Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5838649Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5839275Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5839925Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5840596Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5841239Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5841369Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.5841441Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5841482Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5841518Z unimplemented []
2025-12-04T09:45:15.5841581Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5841683Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5842259Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5842296Z graph_break []
2025-12-04T09:45:15.5842370Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5842409Z Autotune Choices Stats:
2025-12-04T09:45:15.5843150Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.5843279Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5843392Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5843551Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5844236Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5844841Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5845456Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5846062Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5846664Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5847264Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5847869Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5848493Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5850604Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5851228Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5851357Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.5851396Z Autotune Choices Stats:
2025-12-04T09:45:15.5852197Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.5852418Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5852584Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5852868Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5853502Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5854150Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5854785Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5855469Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5856094Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5856725Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5857351Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5857981Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5858624Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5859270Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5859417Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.5859490Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5859533Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5859569Z unimplemented []
2025-12-04T09:45:15.5859629Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5859728Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5860310Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5860351Z graph_break []
2025-12-04T09:45:15.5860454Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5860495Z Autotune Choices Stats:
2025-12-04T09:45:15.5861240Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.5861372Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5861487Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5861647Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5862267Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5862887Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5863508Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5864120Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5864724Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5865329Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5865935Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5866541Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5867168Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5867779Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5867918Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.5867957Z Autotune Choices Stats:
2025-12-04T09:45:15.5868710Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.5868933Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5869099Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5869376Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5870012Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5870668Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5871308Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5871954Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5872594Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5873224Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5873845Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5874476Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5875121Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5875768Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5875895Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.5875968Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5876008Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5876047Z unimplemented []
2025-12-04T09:45:15.5876122Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5876233Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5876806Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5876843Z graph_break []
2025-12-04T09:45:15.5876916Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5876954Z Autotune Choices Stats:
2025-12-04T09:45:15.5877701Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.5877829Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5877943Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5878104Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5878719Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5879334Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5879955Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5880605Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5881221Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5881823Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5882432Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5883047Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5883651Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5884281Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5884411Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.5884450Z Autotune Choices Stats:
2025-12-04T09:45:15.5885226Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.5885455Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5885621Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5885902Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5886531Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5887153Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5887774Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5888417Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5889058Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5889689Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5890311Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5890973Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5891594Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5892220Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5892380Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.5892452Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5892494Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5892530Z unimplemented []
2025-12-04T09:45:15.5892591Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5892690Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5893283Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5893335Z graph_break []
2025-12-04T09:45:15.5893406Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5893447Z Autotune Choices Stats:
2025-12-04T09:45:15.5894177Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.5894307Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5894420Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5894588Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5895206Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5895812Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5896424Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5897039Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5897646Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5898255Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5898864Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5899485Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5900086Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5900746Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5900891Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.5900931Z Autotune Choices Stats:
2025-12-04T09:45:15.5901692Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.5901935Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5902100Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5902378Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5903015Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5903636Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5904274Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5904895Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5905539Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5906175Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5906806Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5907434Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5908053Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5908676Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5908803Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.5908876Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5908916Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5908953Z unimplemented []
2025-12-04T09:45:15.5909012Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5909143Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5909715Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5909752Z graph_break []
2025-12-04T09:45:15.5909823Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5909862Z Autotune Choices Stats:
2025-12-04T09:45:15.5910675Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.5910815Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5910935Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5911095Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5911711Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5912318Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5912944Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5913574Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5914191Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5914816Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5915434Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5916039Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5916643Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5917251Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5917378Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.5917418Z Autotune Choices Stats:
2025-12-04T09:45:15.5918189Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.5918415Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5918581Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5918879Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5919505Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5920130Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5920783Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5921414Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5922061Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5922697Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5923343Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5923987Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5924687Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5925312Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5925442Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.5925514Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5925556Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5925592Z unimplemented []
2025-12-04T09:45:15.5925650Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5925749Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5926335Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5926386Z graph_break []
2025-12-04T09:45:15.5926459Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5926497Z Autotune Choices Stats:
2025-12-04T09:45:15.5927249Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.5927397Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5927510Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5927670Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5928278Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5928882Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5929485Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5930089Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5930747Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5931353Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5931965Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5932588Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5933196Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5933793Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5933925Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.5933963Z Autotune Choices Stats:
2025-12-04T09:45:15.5934713Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.5934950Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5935124Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5935401Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5936039Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5936679Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5937298Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5937920Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5938549Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5939184Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5939815Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5940502Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5941144Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5941770Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5941897Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.5941970Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5942010Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5942048Z unimplemented []
2025-12-04T09:45:15.5942106Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5942208Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5942785Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.5942822Z graph_break []
2025-12-04T09:45:15.5942894Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5942935Z Autotune Choices Stats:
2025-12-04T09:45:15.5943703Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.5943843Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5943959Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5944117Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5944753Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5945357Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5945959Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5946562Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5947158Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5947769Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5948384Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5949004Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5949617Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5950231Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5950357Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.5950397Z Autotune Choices Stats:
2025-12-04T09:45:15.5951197Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.5951414Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5951580Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5951880Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5952519Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5953173Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5953809Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5954433Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5955067Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5955692Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5956339Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5956979Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5957614Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5958240Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5958371Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.5958443Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5958483Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5958518Z unimplemented []
2025-12-04T09:45:15.5958578Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5958675Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5959243Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5959282Z graph_break []
2025-12-04T09:45:15.5959354Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5959394Z Autotune Choices Stats:
2025-12-04T09:45:15.5960133Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.5960279Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5960393Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5960589Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5961206Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5961840Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5962444Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5963050Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5963651Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5964256Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5964881Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5965491Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5966111Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5966713Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5966843Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.5966882Z Autotune Choices Stats:
2025-12-04T09:45:15.5967639Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.5967859Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5968024Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5968300Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5968953Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5969584Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5970233Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5970897Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5971527Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5972154Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5972777Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5973436Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5974070Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5974721Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5974850Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.5974925Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5974964Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5975002Z unimplemented []
2025-12-04T09:45:15.5975060Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5975163Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5975736Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5975775Z graph_break []
2025-12-04T09:45:15.5975846Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5975886Z Autotune Choices Stats:
2025-12-04T09:45:15.5976646Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.5976775Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5976890Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5977049Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5977675Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5978291Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5978906Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5979531Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5980152Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5980785Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5981393Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5982007Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5982627Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5983260Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5983388Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.5983428Z Autotune Choices Stats:
2025-12-04T09:45:15.5984189Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.5984409Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5984576Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5984856Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5985488Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5986127Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5986758Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5987409Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5988035Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5988654Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5989280Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5989903Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5990585Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5991220Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5991379Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.5991452Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.5991494Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.5991530Z unimplemented []
2025-12-04T09:45:15.5991589Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.5991687Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.5992266Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.5992302Z graph_break []
2025-12-04T09:45:15.5992375Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.5992413Z Autotune Choices Stats:
2025-12-04T09:45:15.5993149Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.5993279Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.5993393Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.5993551Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.5994154Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5994777Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5995390Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5996008Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5996612Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5997213Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.5997819Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5998424Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5999042Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5999671Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.5999811Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:15.5999849Z Autotune Choices Stats:
2025-12-04T09:45:15.6000654Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.6000875Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6001041Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6001316Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6001941Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6013759Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6014430Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6015072Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6015748Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6016381Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6017008Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6017642Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6018272Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6018912Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6019057Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:15.6019138Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6019183Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6019221Z unimplemented []
2025-12-04T09:45:15.6019285Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6019414Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6019994Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6020032Z graph_break []
2025-12-04T09:45:15.6020107Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6020148Z Autotune Choices Stats:
2025-12-04T09:45:15.6020925Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.6021056Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6021173Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6021337Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6021955Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6022555Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6023184Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6023806Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6024420Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6025026Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6025648Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6026276Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6026882Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6027506Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6027638Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:15.6027679Z Autotune Choices Stats:
2025-12-04T09:45:15.6028446Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:15.6028679Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6028849Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6029130Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6029769Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6030439Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6031062Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6031702Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6032357Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6032999Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6033621Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6034256Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6034883Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6035521Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6035672Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:15.6035767Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.6035813Z Traceback (most recent call last):
2025-12-04T09:45:15.6035971Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.6036010Z     self.assertTrue(
2025-12-04T09:45:15.6036115Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.6036163Z     raise self.failureException(msg)
2025-12-04T09:45:15.6036292Z AssertionError: False is not true : Log file /tmp/tmpx5r9jo87/flex_attention_configs.json was not created
2025-12-04T09:45:15.6036296Z 
2025-12-04T09:45:15.6036370Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.6036558Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.6036561Z 
2025-12-04T09:45:15.6036651Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.6036725Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6036768Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6036804Z unimplemented []
2025-12-04T09:45:15.6036865Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6037447Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.6037548Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6037584Z graph_break []
2025-12-04T09:45:15.6037656Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6038152Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.6038200Z   current_size = base.storage().size()
2025-12-04T09:45:15.6038241Z Autotune Choices Stats:
2025-12-04T09:45:15.6038983Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.6039112Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6039227Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6039388Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6040020Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6040662Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6041281Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6041882Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6042506Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6043108Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6043707Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6044324Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6044941Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6045557Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6045687Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.6045726Z Autotune Choices Stats:
2025-12-04T09:45:15.6046484Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.6046704Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6046870Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6047150Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6047789Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6048423Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6049057Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6049692Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6050321Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6050982Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6051619Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6052243Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6052885Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6053519Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6053675Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.6053751Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6053792Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6053829Z unimplemented []
2025-12-04T09:45:15.6053888Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6053987Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6054560Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6054600Z graph_break []
2025-12-04T09:45:15.6054672Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6054712Z Autotune Choices Stats:
2025-12-04T09:45:15.6055459Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.6055588Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6055703Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6055861Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6056475Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6057098Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6057713Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6058326Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6058921Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6059525Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6060133Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6060775Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6061394Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6062004Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6062156Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.6062197Z Autotune Choices Stats:
2025-12-04T09:45:15.6062960Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.6063181Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6063348Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6063624Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6064260Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6064886Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6065523Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6066145Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6066796Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6067421Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6068041Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6068667Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6069295Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6069934Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6070071Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.6070143Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6070184Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6070220Z unimplemented []
2025-12-04T09:45:15.6070280Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6070390Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6071057Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6071093Z graph_break []
2025-12-04T09:45:15.6071165Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6071204Z Autotune Choices Stats:
2025-12-04T09:45:15.6071938Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.6072065Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6072178Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6072340Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6072951Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6073569Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6074310Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6074918Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6075531Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6076148Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6076753Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6077356Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6077964Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6078582Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6078721Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.6078761Z Autotune Choices Stats:
2025-12-04T09:45:15.6079531Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.6079769Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6079933Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6080214Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6080888Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6081511Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6082136Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6082796Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6083450Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6084083Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6084715Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6085338Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6085963Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6086585Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6086741Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.6086816Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6086857Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6086894Z unimplemented []
2025-12-04T09:45:15.6086953Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6087053Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6087628Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6087688Z graph_break []
2025-12-04T09:45:15.6087760Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6087799Z Autotune Choices Stats:
2025-12-04T09:45:15.6088546Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.6088675Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6088792Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6088950Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6089571Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6090177Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6090874Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6091494Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6092122Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6092784Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6093389Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6093994Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6094597Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6095207Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6095467Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.6095509Z Autotune Choices Stats:
2025-12-04T09:45:15.6096269Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.6096528Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6096726Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6097008Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6097644Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6098275Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6098898Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6099523Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6100253Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6100981Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6101653Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6102280Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6102913Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6103554Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6103685Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.6103757Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6103797Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6103833Z unimplemented []
2025-12-04T09:45:15.6103892Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6103991Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6104641Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6104677Z graph_break []
2025-12-04T09:45:15.6104749Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6104788Z Autotune Choices Stats:
2025-12-04T09:45:15.6105566Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.6105705Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6105819Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6105979Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6106597Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6107202Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6107812Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6108427Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6109040Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6109654Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6110272Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6110916Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6111517Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6112122Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6112253Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.6112292Z Autotune Choices Stats:
2025-12-04T09:45:15.6113073Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.6113319Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6113486Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6113788Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6114435Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6115056Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6115681Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6116302Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6116931Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6117587Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6118213Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6118861Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6119494Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6120111Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6120242Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.6120317Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6120357Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6120395Z unimplemented []
2025-12-04T09:45:15.6120489Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6120589Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6121167Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6121223Z graph_break []
2025-12-04T09:45:15.6121308Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6121349Z Autotune Choices Stats:
2025-12-04T09:45:15.6122090Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.6122249Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6122366Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6122526Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6123143Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6123749Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6124348Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6124955Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6125569Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6126184Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6126800Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6127416Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6128040Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6128644Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6128775Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.6128815Z Autotune Choices Stats:
2025-12-04T09:45:15.6129588Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.6129833Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6130009Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6130295Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6130978Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6131614Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6132234Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6132860Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6133489Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6134125Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6134760Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6135401Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6136030Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6136656Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6136785Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.6136859Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6136898Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6136937Z unimplemented []
2025-12-04T09:45:15.6136995Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6137096Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6137673Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6137708Z graph_break []
2025-12-04T09:45:15.6137781Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6137821Z Autotune Choices Stats:
2025-12-04T09:45:15.6138580Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.6138721Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6138836Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6138999Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6139640Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6140254Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6140909Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6141511Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6142111Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6142723Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6143348Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6143967Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6144591Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6145194Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6145324Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.6145363Z Autotune Choices Stats:
2025-12-04T09:45:15.6146134Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.6146352Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6146521Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6146811Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6147451Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6148083Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6148712Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6149439Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6150062Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6150739Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6151381Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6152019Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6152678Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6153314Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6153444Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.6153518Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6153560Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6153596Z unimplemented []
2025-12-04T09:45:15.6153656Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6153753Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6154327Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6154367Z graph_break []
2025-12-04T09:45:15.6154442Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6154483Z Autotune Choices Stats:
2025-12-04T09:45:15.6155229Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.6155359Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6155496Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6155656Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6156279Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6156908Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6157516Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6158116Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6158718Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6159323Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6159937Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6160589Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6161213Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6161835Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6161965Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.6162006Z Autotune Choices Stats:
2025-12-04T09:45:15.6162761Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.6162979Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6163148Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6163425Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6164080Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6164712Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6165347Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6165979Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6166606Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6167238Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6167866Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6168518Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6169155Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6169794Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6169936Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.6170009Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6170048Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6170086Z unimplemented []
2025-12-04T09:45:15.6170144Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6170243Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6170861Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6170896Z graph_break []
2025-12-04T09:45:15.6170969Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6171008Z Autotune Choices Stats:
2025-12-04T09:45:15.6171755Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.6171882Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6171997Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6172160Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6172792Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6173410Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6174040Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6174640Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6175262Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6175867Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6176475Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6177110Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6177733Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6178364Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6178493Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.6178533Z Autotune Choices Stats:
2025-12-04T09:45:15.6179297Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.6179515Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6179682Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6179964Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6180631Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6181265Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6181902Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6182540Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6183177Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6183809Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6184430Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6185072Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6185723Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6186362Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6186511Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.6186586Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6186628Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6186664Z unimplemented []
2025-12-04T09:45:15.6186725Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6186823Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6187393Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6187432Z graph_break []
2025-12-04T09:45:15.6187505Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6187546Z Autotune Choices Stats:
2025-12-04T09:45:15.6188291Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.6188420Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6188537Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6188699Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6189309Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6189918Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6190589Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6191205Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6191810Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6192426Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6193053Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6193673Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6194320Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6194934Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6195085Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.6195126Z Autotune Choices Stats:
2025-12-04T09:45:15.6195888Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.6196108Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6196277Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6196550Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6197184Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6197806Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6198515Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6199144Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6199802Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6200470Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6201094Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6201741Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6202372Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6203012Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6203152Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.6203227Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6203268Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6203304Z unimplemented []
2025-12-04T09:45:15.6203363Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6203461Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6204067Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6204104Z graph_break []
2025-12-04T09:45:15.6204177Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6204215Z Autotune Choices Stats:
2025-12-04T09:45:15.6204962Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.6205088Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6205202Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6205361Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6205971Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6206575Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6207203Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6207819Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6208441Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6209037Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6209646Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6210251Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6210888Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6211502Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6211643Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.6211681Z Autotune Choices Stats:
2025-12-04T09:45:15.6212458Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.6212689Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6212854Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6213133Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6213768Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6214393Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6215012Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6215648Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6216281Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6216921Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6217543Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6218178Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6218802Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6219421Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6219550Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.6219647Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6219689Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6219725Z unimplemented []
2025-12-04T09:45:15.6219784Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6219882Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6220495Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6220547Z graph_break []
2025-12-04T09:45:15.6220632Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6220674Z Autotune Choices Stats:
2025-12-04T09:45:15.6221415Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.6221542Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6221659Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6221819Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6222433Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6223035Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6223644Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6224277Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6224897Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6225507Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6226119Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6226723Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6227328Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6227934Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6228075Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.6228126Z Autotune Choices Stats:
2025-12-04T09:45:15.6228877Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.6229104Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6229284Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6229562Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6230197Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6230895Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6231524Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6232148Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6232796Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6233448Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6234077Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6234700Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6235326Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6235949Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6236081Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.6236156Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6236197Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6236234Z unimplemented []
2025-12-04T09:45:15.6236293Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6236392Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6236994Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6237031Z graph_break []
2025-12-04T09:45:15.6237103Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6237143Z Autotune Choices Stats:
2025-12-04T09:45:15.6237895Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.6238031Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6238146Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6238304Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6238913Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6239516Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6240134Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6240787Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6241405Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6242021Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6242643Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6243244Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6243850Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6244456Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6244584Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:15.6244623Z Autotune Choices Stats:
2025-12-04T09:45:15.6245416Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.6245644Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6245810Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6246100Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6246742Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6247370Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6247994Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6248633Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6249278Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6249927Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6250603Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6251250Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6251892Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6252516Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6252648Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:15.6252722Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6252762Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6252798Z unimplemented []
2025-12-04T09:45:15.6252857Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6252954Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6253532Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6253580Z graph_break []
2025-12-04T09:45:15.6253665Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6253704Z Autotune Choices Stats:
2025-12-04T09:45:15.6254443Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.6254568Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6254705Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6254867Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6255480Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6256084Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6256691Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6257297Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6257911Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6258523Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6259134Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6259748Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6260355Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6260984Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6261114Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:15.6261157Z Autotune Choices Stats:
2025-12-04T09:45:15.6261918Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:15.6262137Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6262331Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6262607Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6263257Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6263893Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6264536Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6265161Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6265788Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6266430Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6267066Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6267704Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6268334Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6268967Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6269095Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:15.6269169Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6269209Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6269246Z unimplemented []
2025-12-04T09:45:15.6269304Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6269405Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6269984Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6270022Z graph_break []
2025-12-04T09:45:15.6270093Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6270132Z Autotune Choices Stats:
2025-12-04T09:45:15.6271010Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:15.6271148Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6271263Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6271420Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6272044Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6272674Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6273281Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6273878Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6274487Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6275101Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6275716Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6276335Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6276949Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6277566Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6277694Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:15.6277734Z Autotune Choices Stats:
2025-12-04T09:45:15.6278495Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:15.6278712Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6278877Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6279157Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6279813Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6280485Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6281119Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6281761Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6282388Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6283013Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6283646Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6284288Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6284928Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6285560Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6285690Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:15.6285784Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.6285831Z Traceback (most recent call last):
2025-12-04T09:45:15.6285985Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.6286026Z     self.assertTrue(
2025-12-04T09:45:15.6286130Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.6286178Z     raise self.failureException(msg)
2025-12-04T09:45:15.6286305Z AssertionError: False is not true : Log file /tmp/tmpsuzkclcu/flex_attention_configs.json was not created
2025-12-04T09:45:15.6286307Z 
2025-12-04T09:45:15.6286384Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.6286550Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.6286553Z 
2025-12-04T09:45:15.6286643Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.6286716Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6286758Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6286794Z unimplemented []
2025-12-04T09:45:15.6286855Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6287437Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.6287535Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6287593Z graph_break []
2025-12-04T09:45:15.6287665Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6288157Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.6288204Z   current_size = base.storage().size()
2025-12-04T09:45:15.6288245Z Autotune Choices Stats:
2025-12-04T09:45:15.6289006Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.6289146Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6289260Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6289418Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6290039Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6290671Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6291273Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6291893Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6292508Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6293122Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6293734Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6294345Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6294950Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6295566Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6295695Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.6295736Z Autotune Choices Stats:
2025-12-04T09:45:15.6296508Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.6296738Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6296904Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6297190Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6297824Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6298449Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6299070Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6299690Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6300315Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6301000Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6301635Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6302276Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6302904Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6303527Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6303653Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.6303730Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6303773Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6303810Z unimplemented []
2025-12-04T09:45:15.6303871Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6303970Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6304548Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6304600Z graph_break []
2025-12-04T09:45:15.6304693Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6304732Z Autotune Choices Stats:
2025-12-04T09:45:15.6305471Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.6305597Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6305734Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6305895Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6306506Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6307109Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6307730Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6308351Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6308978Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6309588Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6310221Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6310878Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6311478Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6312083Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6312213Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.6312254Z Autotune Choices Stats:
2025-12-04T09:45:15.6313012Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.6313228Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6313425Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6313703Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6314348Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6314983Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6315604Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6316224Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6316846Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6317473Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6318112Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6318750Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6319387Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6320016Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6320145Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.6320218Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6320259Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6320294Z unimplemented []
2025-12-04T09:45:15.6320356Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6320489Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6321063Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6321101Z graph_break []
2025-12-04T09:45:15.6321173Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6321213Z Autotune Choices Stats:
2025-12-04T09:45:15.6321984Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.6322127Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6322242Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6322402Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6325800Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6326420Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6327027Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6327629Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6328235Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6328876Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6329491Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6330108Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6330757Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6331365Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6331492Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.6331532Z Autotune Choices Stats:
2025-12-04T09:45:15.6332365Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.6332585Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6332750Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6333026Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6333683Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6334330Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6334967Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6335595Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6336226Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6336860Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6337503Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6338187Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6338820Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6339469Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6339595Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.6339672Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6339712Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6339750Z unimplemented []
2025-12-04T09:45:15.6339809Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6339908Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6340533Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6340569Z graph_break []
2025-12-04T09:45:15.6340644Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6340683Z Autotune Choices Stats:
2025-12-04T09:45:15.6341435Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.6341562Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6341700Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6341873Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6342555Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6343172Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6343793Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6344416Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6345038Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6345653Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6346269Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6346884Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6347520Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6348130Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6348259Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.6348302Z Autotune Choices Stats:
2025-12-04T09:45:15.6349066Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.6349283Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6349453Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6349731Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6350377Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6351055Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6351704Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6352343Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6352985Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6353614Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6354239Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6354908Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6355553Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6356224Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6356363Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.6356436Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6356478Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6356514Z unimplemented []
2025-12-04T09:45:15.6356573Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6356672Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6357249Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6357287Z graph_break []
2025-12-04T09:45:15.6357359Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6357399Z Autotune Choices Stats:
2025-12-04T09:45:15.6358145Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.6358275Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6358390Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6358547Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6359167Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6359781Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6360401Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6361045Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6361651Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6362252Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6362861Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6363492Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6364109Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6364724Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6364866Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.6364906Z Autotune Choices Stats:
2025-12-04T09:45:15.6365672Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.6365892Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6366058Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6366335Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6366966Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6367596Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6368227Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6368860Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6369500Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6370131Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6370783Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6371415Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6372061Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6372703Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6372829Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.6372931Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6372972Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6373010Z unimplemented []
2025-12-04T09:45:15.6373068Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6373168Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6373744Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6373782Z graph_break []
2025-12-04T09:45:15.6373858Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6373898Z Autotune Choices Stats:
2025-12-04T09:45:15.6374640Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.6374765Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6374883Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6375045Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6375654Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6376267Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6376880Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6377497Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6378099Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6378723Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6379334Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6379935Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6380601Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6381223Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6381353Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.6381429Z Autotune Choices Stats:
2025-12-04T09:45:15.6382181Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.6382399Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6382567Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6382848Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6383491Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6384117Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6384755Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6385387Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6386027Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6386662Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6387286Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6387934Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6388559Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6389201Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6389341Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.6389413Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6389455Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6389491Z unimplemented []
2025-12-04T09:45:15.6389551Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6389649Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6390243Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6390284Z graph_break []
2025-12-04T09:45:15.6390356Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6390398Z Autotune Choices Stats:
2025-12-04T09:45:15.6391174Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.6391304Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6391417Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6391579Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6392205Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6392813Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6393444Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6394063Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6394690Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6395293Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6395906Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6396513Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6397115Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6397735Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6397877Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.6397922Z Autotune Choices Stats:
2025-12-04T09:45:15.6398692Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.6398923Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6399088Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6399371Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6400016Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6400681Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6401308Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6401952Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6402596Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6403264Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6403898Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6404523Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6405150Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6405776Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6405907Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.6406010Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6406052Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6406095Z unimplemented []
2025-12-04T09:45:15.6406155Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6406260Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6406836Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6406889Z graph_break []
2025-12-04T09:45:15.6406974Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6407016Z Autotune Choices Stats:
2025-12-04T09:45:15.6407758Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.6407885Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6408006Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6408168Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6408791Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6409394Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6410012Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6410680Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6411310Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6411932Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6412546Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6413156Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6413758Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6414376Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6414506Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.6414565Z Autotune Choices Stats:
2025-12-04T09:45:15.6415344Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.6415563Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6415766Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6416048Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6416681Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6417327Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6417952Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6418576Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6419224Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6419862Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6420553Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6421185Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6421816Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6422455Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6422589Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.6422661Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6422707Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6422746Z unimplemented []
2025-12-04T09:45:15.6422812Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6422911Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6423518Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6423568Z graph_break []
2025-12-04T09:45:15.6423646Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6423687Z Autotune Choices Stats:
2025-12-04T09:45:15.6424449Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.6424593Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6424708Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6424875Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6425492Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6426094Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6426704Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6427315Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6427944Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6428559Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6429175Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6429781Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6430391Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6431032Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6431165Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.6431205Z Autotune Choices Stats:
2025-12-04T09:45:15.6431984Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.6432220Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6432387Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6432675Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6433338Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6433964Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6434597Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6435246Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6435876Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6436517Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6437167Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6437808Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6438444Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6439069Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6439201Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.6439281Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6439325Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6439367Z unimplemented []
2025-12-04T09:45:15.6439429Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6439533Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6440108Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6440152Z graph_break []
2025-12-04T09:45:15.6440227Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6440293Z Autotune Choices Stats:
2025-12-04T09:45:15.6441075Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.6441203Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6441360Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6441521Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6442138Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6442747Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6443356Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6443960Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6444573Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6445214Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6445836Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6446466Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6447089Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6447697Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6447825Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.6447871Z Autotune Choices Stats:
2025-12-04T09:45:15.6448636Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.6448859Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6449054Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6449341Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6449988Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6450665Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6451302Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6451944Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6452581Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6453215Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6453890Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6454532Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6455182Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6455830Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6455964Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.6456041Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6456088Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6456128Z unimplemented []
2025-12-04T09:45:15.6456192Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6456291Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6456868Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6456909Z graph_break []
2025-12-04T09:45:15.6456988Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6457028Z Autotune Choices Stats:
2025-12-04T09:45:15.6457791Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.6457933Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6458050Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6458215Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6458856Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6459476Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6460087Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6460757Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6461365Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6461993Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6462630Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6463261Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6463880Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6464489Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6464623Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.6464664Z Autotune Choices Stats:
2025-12-04T09:45:15.6465432Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.6465655Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6465820Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6466100Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6466756Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6467398Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6468039Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6468672Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6469325Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6469972Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6470656Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6471322Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6471966Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6472605Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6472738Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.6472819Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6472863Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6472906Z unimplemented []
2025-12-04T09:45:15.6472967Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6473071Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6473650Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6473692Z graph_break []
2025-12-04T09:45:15.6473767Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6473816Z Autotune Choices Stats:
2025-12-04T09:45:15.6474568Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.6474700Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6474838Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6475010Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6475630Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6476252Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6476871Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6477487Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6478103Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6478727Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6479354Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6479971Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6480828Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6481471Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6481602Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.6481649Z Autotune Choices Stats:
2025-12-04T09:45:15.6482413Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.6482638Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6482809Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6483094Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6483733Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6484406Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6485063Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6485696Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6486332Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6486963Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6487592Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6488240Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6488879Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6489518Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6489662Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.6489737Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6489784Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6489824Z unimplemented []
2025-12-04T09:45:15.6489890Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6489992Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6490607Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6490647Z graph_break []
2025-12-04T09:45:15.6490725Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6490766Z Autotune Choices Stats:
2025-12-04T09:45:15.6491516Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.6491651Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6491767Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6491934Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6492571Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6493188Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6493821Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6494443Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6495134Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6495742Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6496368Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6497042Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6497655Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6498279Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6498434Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:15.6498476Z Autotune Choices Stats:
2025-12-04T09:45:15.6499241Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.6499465Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6499632Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6499914Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6500632Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6501295Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6501954Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6502674Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6503352Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6504010Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6504654Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6505298Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6505979Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6506643Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6506782Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:15.6506905Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6506956Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6507036Z unimplemented []
2025-12-04T09:45:15.6507105Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6507235Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6507826Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6507883Z graph_break []
2025-12-04T09:45:15.6507996Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6508049Z Autotune Choices Stats:
2025-12-04T09:45:15.6508819Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.6508961Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6509097Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6509298Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6509919Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6510608Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6511253Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6511891Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6512539Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6513174Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6513803Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6514419Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6515062Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6515709Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6515863Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:15.6515926Z Autotune Choices Stats:
2025-12-04T09:45:15.6516716Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:15.6516975Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6517156Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6517460Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6518104Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6518758Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6519442Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6520102Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6520802Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6521467Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6524189Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6524826Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6525455Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6526132Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6526283Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:15.6526361Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6526405Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6526445Z unimplemented []
2025-12-04T09:45:15.6526506Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6526609Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6527207Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6527256Z graph_break []
2025-12-04T09:45:15.6527332Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6527372Z Autotune Choices Stats:
2025-12-04T09:45:15.6528121Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:15.6528252Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6528368Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6528531Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6529148Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6529755Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6530378Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6531028Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6531660Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6532259Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6532867Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6533472Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6534086Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6534721Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6534864Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:15.6534903Z Autotune Choices Stats:
2025-12-04T09:45:15.6535669Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:15.6535903Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6536070Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6536348Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6536981Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6537608Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6538235Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6538878Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6539512Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6540154Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6540825Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6541457Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6542087Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6542711Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6542843Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:15.6542917Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6542996Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6543033Z unimplemented []
2025-12-04T09:45:15.6543095Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6543195Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6543768Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6543805Z graph_break []
2025-12-04T09:45:15.6543894Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6543948Z Autotune Choices Stats:
2025-12-04T09:45:15.6544694Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:15.6544823Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6544938Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6545101Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6545710Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6546314Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6546920Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6547537Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6548148Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6548780Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6549386Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6550009Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6550655Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6551254Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6551382Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:15.6551424Z Autotune Choices Stats:
2025-12-04T09:45:15.6552210Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.6552429Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6552622Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6552898Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6553533Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6554161Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6554783Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6555406Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6556062Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6556699Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6557341Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6557969Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6558598Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6559241Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6559368Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:15.6559462Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.6559508Z Traceback (most recent call last):
2025-12-04T09:45:15.6559665Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.6559703Z     self.assertTrue(
2025-12-04T09:45:15.6559810Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.6559858Z     raise self.failureException(msg)
2025-12-04T09:45:15.6560000Z AssertionError: False is not true : Log file /tmp/tmpvzrmqh1r/flex_attention_configs.json was not created
2025-12-04T09:45:15.6560011Z 
2025-12-04T09:45:15.6560088Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.6560252Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.6560254Z 
2025-12-04T09:45:15.6560343Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.6560449Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6560494Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6560531Z unimplemented []
2025-12-04T09:45:15.6560593Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6561193Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.6561309Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6561345Z graph_break []
2025-12-04T09:45:15.6561417Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6561911Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.6561962Z   current_size = base.storage().size()
2025-12-04T09:45:15.6562005Z Autotune Choices Stats:
2025-12-04T09:45:15.6562744Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.6562873Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6562989Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6563152Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6563764Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6564378Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6564993Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6565622Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6566226Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6566827Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6567449Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6568050Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6568658Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6569269Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6569399Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.6569454Z Autotune Choices Stats:
2025-12-04T09:45:15.6570229Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.6570489Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6570656Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6570938Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6571574Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6572218Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6572851Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6573486Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6574127Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6574760Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6575381Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6576002Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6576623Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6577253Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6577393Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.6577466Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6577508Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6577545Z unimplemented []
2025-12-04T09:45:15.6577604Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6577703Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6578288Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6578335Z graph_break []
2025-12-04T09:45:15.6578407Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6578444Z Autotune Choices Stats:
2025-12-04T09:45:15.6579185Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.6579313Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6579428Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6579588Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6580208Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6580948Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6581570Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6582191Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6582828Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6583430Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6584036Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6584645Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6585243Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6585860Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6585996Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.6586036Z Autotune Choices Stats:
2025-12-04T09:45:15.6586804Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.6587032Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6587197Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6587480Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6588113Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6588745Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6589371Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6590008Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6590685Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6591330Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6591964Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6592594Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6593218Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6593835Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6593963Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.6594036Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6594098Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6594146Z unimplemented []
2025-12-04T09:45:15.6594207Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6594309Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6594885Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6594921Z graph_break []
2025-12-04T09:45:15.6594995Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6595044Z Autotune Choices Stats:
2025-12-04T09:45:15.6595798Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.6595925Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6596038Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6596203Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6596820Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6598996Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6600982Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6602994Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6604214Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6605128Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6605918Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6606705Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6607513Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6608291Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6608473Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.6608536Z Autotune Choices Stats:
2025-12-04T09:45:15.6609571Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.6609863Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6610102Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6610542Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6611388Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6612204Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6613012Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6613774Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6614437Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6615094Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6615783Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6616434Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6617082Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6617726Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6617870Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.6617976Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6618023Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6618069Z unimplemented []
2025-12-04T09:45:15.6618136Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6618248Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6618881Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6618939Z graph_break []
2025-12-04T09:45:15.6619019Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6619067Z Autotune Choices Stats:
2025-12-04T09:45:15.6619853Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.6620006Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6620130Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6620297Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6620974Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6621599Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6622222Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6622846Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6623497Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6624116Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6624743Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6625341Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6625946Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6626554Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6626685Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.6626730Z Autotune Choices Stats:
2025-12-04T09:45:15.6627501Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.6627733Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6627906Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6628191Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6628845Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6629464Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6630084Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6630734Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6631359Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6632001Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6632644Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6633299Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6633923Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6634543Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6634676Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.6634752Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6634801Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6634841Z unimplemented []
2025-12-04T09:45:15.6634906Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6635010Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6635582Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6635620Z graph_break []
2025-12-04T09:45:15.6635698Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6635741Z Autotune Choices Stats:
2025-12-04T09:45:15.6636505Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.6636639Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6636756Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6636943Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6637550Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6638159Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6638770Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6639372Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6639969Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6640640Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6641258Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6641877Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6642479Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6643086Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6643218Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.6643262Z Autotune Choices Stats:
2025-12-04T09:45:15.6644015Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.6644238Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6644421Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6644715Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6645345Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6645989Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6646613Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6647233Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6647861Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6648485Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6649131Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6649785Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6650454Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6651076Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6651211Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.6651290Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6651334Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6651377Z unimplemented []
2025-12-04T09:45:15.6651439Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6651543Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6652118Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6652163Z graph_break []
2025-12-04T09:45:15.6652236Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6652281Z Autotune Choices Stats:
2025-12-04T09:45:15.6653036Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.6653183Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6653319Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6653481Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6654105Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6654725Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6655323Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6655921Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6656525Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6657129Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6657755Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6658364Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6658984Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6659585Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6659716Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.6659760Z Autotune Choices Stats:
2025-12-04T09:45:15.6660562Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.6660786Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6660955Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6661231Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6661875Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6662527Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6663164Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6663782Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6664403Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6665038Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6665664Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6666319Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6666948Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6667586Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6667718Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.6667791Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6667839Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6667879Z unimplemented []
2025-12-04T09:45:15.6667943Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6668045Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6668620Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6668657Z graph_break []
2025-12-04T09:45:15.6668735Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6668777Z Autotune Choices Stats:
2025-12-04T09:45:15.6669519Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.6669649Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6669764Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6669951Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6670606Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6671228Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6671846Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6672446Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6673049Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6673651Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6674269Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6674887Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6675498Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6676108Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6676241Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.6676281Z Autotune Choices Stats:
2025-12-04T09:45:15.6677042Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.6677264Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6677433Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6677713Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6678337Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6678981Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6679612Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6680241Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6680900Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6681525Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6682149Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6682774Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6683434Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6684065Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6684210Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.6684284Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6684330Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6684371Z unimplemented []
2025-12-04T09:45:15.6684433Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6684536Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6685111Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6685154Z graph_break []
2025-12-04T09:45:15.6685228Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6685272Z Autotune Choices Stats:
2025-12-04T09:45:15.6686003Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.6686135Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6686253Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6686412Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6687034Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6687646Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6688255Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6688868Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6689472Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6690077Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6690711Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6691363Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6691975Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6692589Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6692732Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.6692777Z Autotune Choices Stats:
2025-12-04T09:45:15.6693533Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.6693755Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6693925Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6694201Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6694832Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6695457Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6696097Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6696729Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6697373Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6698004Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6698627Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6699253Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6699890Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6700546Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6700675Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.6700753Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6700815Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6700879Z unimplemented []
2025-12-04T09:45:15.6700941Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6701047Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6701628Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6701666Z graph_break []
2025-12-04T09:45:15.6701743Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6701785Z Autotune Choices Stats:
2025-12-04T09:45:15.6702525Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.6702657Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6702774Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6702942Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6703557Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6704171Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6704785Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6705404Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6706015Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6706616Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6707223Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6707839Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6708469Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6709078Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6709211Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.6709252Z Autotune Choices Stats:
2025-12-04T09:45:15.6710035Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.6710257Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6710448Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6710731Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6711360Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6711990Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6712626Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6713260Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6713908Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6714548Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6715173Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6715808Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6716428Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6717064Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6717208Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.6717282Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6717328Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6717367Z unimplemented []
2025-12-04T09:45:15.6717432Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6717533Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6718129Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6718184Z graph_break []
2025-12-04T09:45:15.6718257Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6718301Z Autotune Choices Stats:
2025-12-04T09:45:15.6719042Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.6719175Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6719295Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6719462Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6720085Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6720730Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6721352Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6721985Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6722615Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6723237Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6723850Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6724461Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6725071Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6725684Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6725829Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.6725873Z Autotune Choices Stats:
2025-12-04T09:45:15.6726653Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.6726889Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6727059Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6727339Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6727976Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6728602Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6729227Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6729866Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6730545Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6731212Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6731862Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6732516Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6733150Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6733779Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6733908Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.6733993Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6734037Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6734096Z unimplemented []
2025-12-04T09:45:15.6734171Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6734277Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6734862Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6734900Z graph_break []
2025-12-04T09:45:15.6734977Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6735031Z Autotune Choices Stats:
2025-12-04T09:45:15.6735785Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.6735915Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6736033Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6736199Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6736812Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6737424Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6738039Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6738658Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6739283Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6739916Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6740594Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6741202Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6741814Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6742424Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6742557Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.6742598Z Autotune Choices Stats:
2025-12-04T09:45:15.6743385Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.6743618Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6743808Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6744104Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6744747Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6745381Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6746018Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6746668Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6747330Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6747988Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6748633Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6749284Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6749937Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6750623Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6750760Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.6750835Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6750882Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6750921Z unimplemented []
2025-12-04T09:45:15.6750985Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6751085Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6751686Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6751742Z graph_break []
2025-12-04T09:45:15.6751817Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6751860Z Autotune Choices Stats:
2025-12-04T09:45:15.6752626Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.6752775Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6752890Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6753057Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6753688Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6754309Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6754939Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6755570Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6756222Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6756855Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6757507Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6758130Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6758751Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6759373Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6759510Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.6759555Z Autotune Choices Stats:
2025-12-04T09:45:15.6760347Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.6760690Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6760864Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6761153Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6761829Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6762491Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6763134Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6763799Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6764478Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6765163Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6765842Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6766541Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6767212Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6767888Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6768024Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.6768106Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6768152Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6768198Z unimplemented []
2025-12-04T09:45:15.6768264Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6768375Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6768984Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6769028Z graph_break []
2025-12-04T09:45:15.6769110Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6769154Z Autotune Choices Stats:
2025-12-04T09:45:15.6769978Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.6770114Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6770240Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6770472Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6771127Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6771777Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6772433Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6773078Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6773732Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6774429Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6775118Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6775790Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6776471Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6777130Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6777277Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:15.6777321Z Autotune Choices Stats:
2025-12-04T09:45:15.6778151Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.6778389Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6778590Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6778906Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6779588Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6780299Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6781021Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6781699Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6782385Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6783075Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6783773Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6784477Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6785189Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6785874Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6786019Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:15.6786101Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6786151Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6786193Z unimplemented []
2025-12-04T09:45:15.6786262Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6786371Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6787013Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6787057Z graph_break []
2025-12-04T09:45:15.6787137Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6787184Z Autotune Choices Stats:
2025-12-04T09:45:15.6788008Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.6788164Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6788290Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6788472Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6789162Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6789831Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6790536Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6791198Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6791862Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6792515Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6793217Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6793896Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6794574Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6795247Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6795391Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:15.6795435Z Autotune Choices Stats:
2025-12-04T09:45:15.6796257Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:15.6796501Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6796682Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6796988Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6797689Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6798386Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6799093Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6799779Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6800500Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6801185Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6801869Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6802581Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6803288Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6803988Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6804128Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:15.6804212Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6804259Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6804305Z unimplemented []
2025-12-04T09:45:15.6804373Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6804488Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6805108Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6805151Z graph_break []
2025-12-04T09:45:15.6805231Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6805279Z Autotune Choices Stats:
2025-12-04T09:45:15.6806088Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:15.6806227Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6806354Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6806543Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6807216Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6807888Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6808563Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6809221Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6809876Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6810587Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6811249Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6811937Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6812606Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6813286Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6813426Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:15.6813474Z Autotune Choices Stats:
2025-12-04T09:45:15.6814303Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:15.6814540Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6814727Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6815030Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6815715Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6816414Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6817121Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6817809Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6818498Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6819186Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6819870Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6820608Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6821324Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6822021Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6822183Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:15.6822264Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6822314Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6822356Z unimplemented []
2025-12-04T09:45:15.6822424Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6822533Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6823158Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6823200Z graph_break []
2025-12-04T09:45:15.6823283Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6823326Z Autotune Choices Stats:
2025-12-04T09:45:15.6824146Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:15.6824291Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6824417Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6824595Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6825274Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6825938Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6826606Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6827282Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6827947Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6828612Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6829279Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6829952Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6830665Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6831340Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6831507Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:15.6831551Z Autotune Choices Stats:
2025-12-04T09:45:15.6832378Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.6832622Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6832802Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6833114Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6833805Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6834489Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6835198Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6835894Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6836588Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6837273Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6837953Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6838639Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6839347Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6840034Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6840176Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:15.6840260Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6840318Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6840363Z unimplemented []
2025-12-04T09:45:15.6840486Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6840600Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6841228Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6841272Z graph_break []
2025-12-04T09:45:15.6841352Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6841399Z Autotune Choices Stats:
2025-12-04T09:45:15.6842211Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:15.6842350Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6842477Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6842652Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6843319Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6844004Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6844675Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6845337Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6846014Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6846672Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6847334Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6848004Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6848679Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6849349Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6849488Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:15.6849535Z Autotune Choices Stats:
2025-12-04T09:45:15.6850378Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:15.6850678Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6850863Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6851167Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6851861Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6852550Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6853278Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6853970Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6854690Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6855388Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6856069Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6856751Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6857443Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6858139Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6858294Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:15.6858394Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.6858450Z Traceback (most recent call last):
2025-12-04T09:45:15.6858624Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.6858671Z     self.assertTrue(
2025-12-04T09:45:15.6858792Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.6858849Z     raise self.failureException(msg)
2025-12-04T09:45:15.6858990Z AssertionError: False is not true : Log file /tmp/tmp6to3xt_d/flex_attention_configs.json was not created
2025-12-04T09:45:15.6859008Z 
2025-12-04T09:45:15.6859109Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.6859290Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.6859293Z 
2025-12-04T09:45:15.6859393Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.6859477Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6859528Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6859569Z unimplemented []
2025-12-04T09:45:15.6859654Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6860287Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.6860401Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6860473Z graph_break []
2025-12-04T09:45:15.6860553Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6861104Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.6861157Z   current_size = base.storage().size()
2025-12-04T09:45:15.6861207Z Autotune Choices Stats:
2025-12-04T09:45:15.6862029Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.6862174Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6862304Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6862510Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6863180Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6863855Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6864521Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6865196Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6865849Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6866506Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6867165Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6867850Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6868517Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6869184Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6869324Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.6869371Z Autotune Choices Stats:
2025-12-04T09:45:15.6870196Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.6870481Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6870667Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6870970Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6871660Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6872369Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6873070Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6873762Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6874452Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6875137Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6875811Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6876505Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6877206Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6877897Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6878048Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.6878134Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6878181Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6878225Z unimplemented []
2025-12-04T09:45:15.6878291Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6878405Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6879039Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6879082Z graph_break []
2025-12-04T09:45:15.6879166Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6879209Z Autotune Choices Stats:
2025-12-04T09:45:15.6880021Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.6880164Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6880290Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6880506Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6881188Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6881858Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6882530Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6883197Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6883857Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6884509Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6885174Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6885854Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6886538Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6887210Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6887367Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.6887410Z Autotune Choices Stats:
2025-12-04T09:45:15.6888239Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.6888486Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6888666Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6888971Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6889664Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6890337Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6891094Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6891794Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6892490Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6893182Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6893868Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6894552Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6895230Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6895933Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6896077Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.6896157Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6896207Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6896260Z unimplemented []
2025-12-04T09:45:15.6896342Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6896453Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6897072Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6897117Z graph_break []
2025-12-04T09:45:15.6897197Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6897244Z Autotune Choices Stats:
2025-12-04T09:45:15.6898051Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.6898194Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6898322Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6898500Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6899174Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6899840Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6900544Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6901213Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6901885Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6902542Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6903190Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6903863Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6904544Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6905206Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6905349Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.6905396Z Autotune Choices Stats:
2025-12-04T09:45:15.6906224Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.6906484Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6906671Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6906978Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6907661Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6908344Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6909024Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6909724Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6910453Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6911145Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6911829Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6912500Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6913186Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6913884Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6914038Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.6914123Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6914169Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6914213Z unimplemented []
2025-12-04T09:45:15.6914279Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6914390Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6915037Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6915091Z graph_break []
2025-12-04T09:45:15.6915173Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6915216Z Autotune Choices Stats:
2025-12-04T09:45:15.6916029Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.6916170Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6916297Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6916475Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6917140Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6917808Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6918484Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6919146Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6919835Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6920538Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6921205Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6921864Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6922520Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6923185Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6923342Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.6923385Z Autotune Choices Stats:
2025-12-04T09:45:15.6924218Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.6924473Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6924657Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6924957Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6925640Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6926320Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6927002Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6927694Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6928386Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6929083Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6929775Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6930491Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6931169Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6931845Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6931987Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.6932066Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6932116Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6932158Z unimplemented []
2025-12-04T09:45:15.6932245Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6932369Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6932993Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6933037Z graph_break []
2025-12-04T09:45:15.6933118Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6933165Z Autotune Choices Stats:
2025-12-04T09:45:15.6934005Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.6934148Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6934273Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6934453Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6935121Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6935786Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6936451Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6937119Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6937801Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6938466Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6939142Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6939806Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6940509Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6941165Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6941309Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.6941356Z Autotune Choices Stats:
2025-12-04T09:45:15.6942201Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.6942457Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6942656Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6942972Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6943663Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6944354Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6945033Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6945716Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6946422Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6947118Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6947802Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6948493Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6949189Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6949869Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6950010Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.6950093Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6950140Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6950184Z unimplemented []
2025-12-04T09:45:15.6950249Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6950360Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6951039Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6951101Z graph_break []
2025-12-04T09:45:15.6951184Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6951227Z Autotune Choices Stats:
2025-12-04T09:45:15.6952051Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.6952207Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6952334Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6952506Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6953180Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6953841Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6954504Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6955159Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6955830Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6956498Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6957182Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6957833Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6958499Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6959154Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6959299Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.6959342Z Autotune Choices Stats:
2025-12-04T09:45:15.6960166Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.6960473Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6960658Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6960964Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6961669Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6962359Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6963041Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6963744Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6964417Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6965104Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6965796Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6966483Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6967172Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6967862Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6968006Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.6968086Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6968135Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6968176Z unimplemented []
2025-12-04T09:45:15.6968247Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6968357Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6968981Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.6969025Z graph_break []
2025-12-04T09:45:15.6969104Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6969151Z Autotune Choices Stats:
2025-12-04T09:45:15.6969978Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.6970131Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6970255Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6970487Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6971177Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6971836Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6972515Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6973186Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6973859Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6974535Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6975210Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6975891Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6976551Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6977215Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6977359Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.6977402Z Autotune Choices Stats:
2025-12-04T09:45:15.6978222Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.6978463Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6978644Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6978972Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6979655Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6980344Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6981090Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6981772Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6982451Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6983126Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6983824Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6984520Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6985231Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6985913Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6986054Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.6986139Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.6986184Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.6986229Z unimplemented []
2025-12-04T09:45:15.6986296Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.6986409Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.6987057Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.6987102Z graph_break []
2025-12-04T09:45:15.6987182Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.6987229Z Autotune Choices Stats:
2025-12-04T09:45:15.6988040Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.6988204Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6988331Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6988504Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6989183Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6989855Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6990575Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6991234Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6991901Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6992566Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.6993249Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6993923Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6994751Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6995408Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6995550Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.6995598Z Autotune Choices Stats:
2025-12-04T09:45:15.6996420Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.6996660Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.6996845Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.6997152Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.6997849Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6998538Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6999242Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.6999915Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7000647Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7001338Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7002022Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7002745Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7003441Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7004153Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7004296Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.7004376Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7004425Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7004467Z unimplemented []
2025-12-04T09:45:15.7004536Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7004650Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7005291Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7005333Z graph_break []
2025-12-04T09:45:15.7005416Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7005460Z Autotune Choices Stats:
2025-12-04T09:45:15.7006268Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.7006410Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7006534Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7006710Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7007410Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7008080Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7008765Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7009425Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7010088Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7010790Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7011453Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7012146Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7012834Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7013503Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7013647Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.7013691Z Autotune Choices Stats:
2025-12-04T09:45:15.7014527Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.7014769Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7014950Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7015261Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7015960Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7016660Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7017351Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7018055Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7018744Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7019428Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7020128Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7020864Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7021566Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7022275Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7022433Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.7022517Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7022562Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7022607Z unimplemented []
2025-12-04T09:45:15.7022673Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7022786Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7023420Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7023465Z graph_break []
2025-12-04T09:45:15.7023545Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7023592Z Autotune Choices Stats:
2025-12-04T09:45:15.7024404Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.7024546Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7024675Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7024850Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7025542Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7026209Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7026878Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7027553Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7028209Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7028869Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7029536Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7030199Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7030957Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7031626Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7031780Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.7031827Z Autotune Choices Stats:
2025-12-04T09:45:15.7032660Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.7032900Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7033082Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7033381Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7034072Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7034758Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7035448Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7036149Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7036843Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7037543Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7038220Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7038903Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7039585Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7040286Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7040448Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.7040531Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7040580Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7040635Z unimplemented []
2025-12-04T09:45:15.7040704Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7040837Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7041464Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7041505Z graph_break []
2025-12-04T09:45:15.7041589Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7041632Z Autotune Choices Stats:
2025-12-04T09:45:15.7042435Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.7042578Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7042703Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7042879Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7043545Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7044234Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7044913Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7045588Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7048285Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7048949Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7049656Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7050319Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7051023Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7051709Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7051852Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.7051900Z Autotune Choices Stats:
2025-12-04T09:45:15.7052746Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.7053058Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7053242Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7053551Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7054243Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7054922Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7055610Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7056309Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7057007Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7057694Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7058393Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7059074Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7059753Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7060494Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7066282Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.7066378Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7066428Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7066469Z unimplemented []
2025-12-04T09:45:15.7066540Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7066653Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7067324Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7067384Z graph_break []
2025-12-04T09:45:15.7067468Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7067532Z Autotune Choices Stats:
2025-12-04T09:45:15.7068346Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.7068491Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7068621Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7068799Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7069478Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7070132Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7070868Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7071528Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7072202Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7073114Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7073800Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7074463Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7075122Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7075802Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7075946Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.7075990Z Autotune Choices Stats:
2025-12-04T09:45:15.7076814Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.7077075Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7077259Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7077575Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7078258Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7078942Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7079618Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7080304Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7081083Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7081774Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7082459Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7083162Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7083842Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7084538Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7084679Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.7084761Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7084808Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7084850Z unimplemented []
2025-12-04T09:45:15.7084916Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7085040Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7085665Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7085707Z graph_break []
2025-12-04T09:45:15.7085789Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7085833Z Autotune Choices Stats:
2025-12-04T09:45:15.7086656Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.7086817Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7086944Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7087116Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7087795Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7088467Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7089136Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7089815Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7090512Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7091195Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7091882Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7092543Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7093215Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7093873Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7094013Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:15.7094057Z Autotune Choices Stats:
2025-12-04T09:45:15.7094894Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.7095132Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7095317Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7095642Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7096327Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7097018Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7097701Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7098386Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7099079Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7099761Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7100497Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7101208Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7101889Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7102566Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7102708Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:15.7102787Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7102834Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7102873Z unimplemented []
2025-12-04T09:45:15.7102939Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7103047Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7103686Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7103726Z graph_break []
2025-12-04T09:45:15.7103806Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7103850Z Autotune Choices Stats:
2025-12-04T09:45:15.7104660Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.7104824Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7104948Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7105132Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7105802Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7106462Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7107128Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7107784Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7108455Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7109104Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7109769Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7110506Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7111168Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7111823Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7111965Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:15.7112006Z Autotune Choices Stats:
2025-12-04T09:45:15.7112834Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:15.7113091Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7113271Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7113571Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7114273Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7114977Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7115657Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7116349Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7117030Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7117747Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7118421Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7119119Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7119822Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7120539Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7120678Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:15.7120757Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7120801Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7120841Z unimplemented []
2025-12-04T09:45:15.7120907Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7121016Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7121637Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7121681Z graph_break []
2025-12-04T09:45:15.7121760Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7121804Z Autotune Choices Stats:
2025-12-04T09:45:15.7122621Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:15.7122760Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7122883Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7123054Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7123747Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7124421Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7125081Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7125740Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7126392Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7127065Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7127722Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7128408Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7129075Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7129729Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7129868Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:15.7129911Z Autotune Choices Stats:
2025-12-04T09:45:15.7130778Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:15.7131015Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7131195Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7131513Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7132198Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7132898Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7133607Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7134294Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7134985Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7135673Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7136391Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7137070Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7137777Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7138466Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7138608Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:15.7138687Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7138734Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7138773Z unimplemented []
2025-12-04T09:45:15.7138840Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7138947Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7139576Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7139616Z graph_break []
2025-12-04T09:45:15.7139696Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7139740Z Autotune Choices Stats:
2025-12-04T09:45:15.7140583Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:15.7140738Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7140862Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7141036Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7141724Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7142394Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7143063Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7143727Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7144405Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7145063Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7145735Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7146400Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7147080Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7147749Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7147888Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:15.7147931Z Autotune Choices Stats:
2025-12-04T09:45:15.7148771Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.7149011Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7149189Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7149493Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7150205Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7150925Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7151629Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7152323Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7153008Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7153689Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7154383Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7155083Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7155764Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7156464Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7156613Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:15.7156692Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7156736Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7156777Z unimplemented []
2025-12-04T09:45:15.7156843Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7156954Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7157573Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7157615Z graph_break []
2025-12-04T09:45:15.7157693Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7157737Z Autotune Choices Stats:
2025-12-04T09:45:15.7158545Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:15.7158683Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7158807Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7158982Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7159666Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7160337Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7161045Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7161717Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7162374Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7163027Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7163686Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7164357Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7165014Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7165691Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7165840Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:15.7165884Z Autotune Choices Stats:
2025-12-04T09:45:15.7166712Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:15.7166949Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7167127Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7167430Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7168127Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7168819Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7169503Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7170209Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7170926Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7171605Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7172280Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7172963Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7173657Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7174349Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7174505Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:15.7174584Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7174643Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7174682Z unimplemented []
2025-12-04T09:45:15.7174747Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7174854Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7175478Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7175518Z graph_break []
2025-12-04T09:45:15.7175597Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7175640Z Autotune Choices Stats:
2025-12-04T09:45:15.7176454Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:15.7176595Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7176718Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7176896Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7177580Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7178246Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7178919Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7179581Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7180243Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7180928Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7181591Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7182252Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7182926Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7183594Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7183749Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:15.7183791Z Autotune Choices Stats:
2025-12-04T09:45:15.7184629Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:15.7184869Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7185047Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7185352Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7186045Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7186743Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7187431Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7188131Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7188821Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7189512Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7190189Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7190914Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7191598Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7192298Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7192439Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:15.7192539Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.7192589Z Traceback (most recent call last):
2025-12-04T09:45:15.7192761Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.7192819Z     self.assertTrue(
2025-12-04T09:45:15.7192947Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.7192999Z     raise self.failureException(msg)
2025-12-04T09:45:15.7193137Z AssertionError: False is not true : Log file /tmp/tmpt6axgtf_/flex_attention_configs.json was not created
2025-12-04T09:45:15.7193154Z 
2025-12-04T09:45:15.7193237Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.7193418Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.7193420Z 
2025-12-04T09:45:15.7193516Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.7193597Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7193641Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7193683Z unimplemented []
2025-12-04T09:45:15.7193749Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7194381Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.7194490Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7194529Z graph_break []
2025-12-04T09:45:15.7194608Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7195152Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.7195205Z   current_size = base.storage().size()
2025-12-04T09:45:15.7195248Z Autotune Choices Stats:
2025-12-04T09:45:15.7196057Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.7196210Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7196337Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7196515Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7197192Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7197857Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7198533Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7199191Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7199847Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7200582Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7201267Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7201927Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7202610Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7203275Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7203417Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.7203460Z Autotune Choices Stats:
2025-12-04T09:45:15.7204310Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.7204549Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7204729Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7205042Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7205751Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7206428Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7207132Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7207827Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7208505Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7209186Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7209872Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7210599Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7211278Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7211974Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7212128Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.7212208Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7212254Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7212294Z unimplemented []
2025-12-04T09:45:15.7212362Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7212469Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7213096Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7213137Z graph_break []
2025-12-04T09:45:15.7213215Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7213260Z Autotune Choices Stats:
2025-12-04T09:45:15.7214065Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.7214205Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7214331Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7214506Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7215179Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7215848Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7216514Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7217180Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7217829Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7218488Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7219146Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7219829Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7220524Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7221537Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7221696Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.7221742Z Autotune Choices Stats:
2025-12-04T09:45:15.7222576Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.7222817Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7222996Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7223302Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7223988Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7224684Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7225358Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7226060Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7226755Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7227437Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7228120Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7228804Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7229501Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7230183Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7230343Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.7230480Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7230539Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7230581Z unimplemented []
2025-12-04T09:45:15.7230645Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7230755Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7231379Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7231418Z graph_break []
2025-12-04T09:45:15.7231498Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7231539Z Autotune Choices Stats:
2025-12-04T09:45:15.7232347Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.7232488Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7232611Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7232787Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7233448Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7234122Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7234801Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7235470Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7236136Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7236810Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7237464Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7238122Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7238803Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7239469Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7239627Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.7239668Z Autotune Choices Stats:
2025-12-04T09:45:15.7240548Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.7240801Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7240985Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7241293Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7241982Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7242660Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7243355Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7244036Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7244743Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7245438Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7246122Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7246802Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7247476Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7248168Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7248311Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.7248390Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7248437Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7248478Z unimplemented []
2025-12-04T09:45:15.7248544Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7248672Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7249296Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7249347Z graph_break []
2025-12-04T09:45:15.7249427Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7249472Z Autotune Choices Stats:
2025-12-04T09:45:15.7250280Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.7250442Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7250565Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7250740Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7251406Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7252073Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7252747Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7253419Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7254090Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7254756Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7255417Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7256072Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7256740Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7257423Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7257564Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.7257608Z Autotune Choices Stats:
2025-12-04T09:45:15.7258456Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.7258714Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7258896Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7259201Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7259888Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7260603Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7261286Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7261976Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7262677Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7263366Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7264054Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7264737Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7265423Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7266117Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7266267Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.7266349Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7266393Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7266434Z unimplemented []
2025-12-04T09:45:15.7266498Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7266606Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7267242Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7267296Z graph_break []
2025-12-04T09:45:15.7267377Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7267418Z Autotune Choices Stats:
2025-12-04T09:45:15.7268237Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.7268376Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7268503Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7268678Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7269339Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7269999Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7270831Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7271485Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7272165Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7272837Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7273511Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7274170Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7274829Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7275483Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7275643Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.7275685Z Autotune Choices Stats:
2025-12-04T09:45:15.7276516Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.7276774Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7276953Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7277263Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7277950Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7278645Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7279323Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7280002Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7280734Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7281428Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7282116Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7282809Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7283489Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7284172Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7284314Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.7284393Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7284438Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7284477Z unimplemented []
2025-12-04T09:45:15.7284543Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7284650Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7285296Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7285338Z graph_break []
2025-12-04T09:45:15.7285418Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7285467Z Autotune Choices Stats:
2025-12-04T09:45:15.7286289Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.7286454Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7286577Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7286751Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7287423Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7288080Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7288742Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7289418Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7290073Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7290778Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7291465Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7292128Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7292807Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7293470Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7293612Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.7293655Z Autotune Choices Stats:
2025-12-04T09:45:15.7294497Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.7294736Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7294915Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7295242Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7295933Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7296620Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7297299Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7297981Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7298686Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7299367Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7300070Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7300806Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7301501Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7302192Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7302333Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.7302414Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7302458Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7302500Z unimplemented []
2025-12-04T09:45:15.7302565Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7302673Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7303311Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7303367Z graph_break []
2025-12-04T09:45:15.7303446Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7303491Z Autotune Choices Stats:
2025-12-04T09:45:15.7304299Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.7304464Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7304591Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7304777Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7305444Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7306104Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7306760Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7307418Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7308082Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7308742Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7309430Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7310106Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7310815Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7311468Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7311610Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.7311653Z Autotune Choices Stats:
2025-12-04T09:45:15.7312481Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.7312736Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7312917Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7313220Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7313926Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7314619Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7315304Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7315980Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7316663Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7317373Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7318049Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7318741Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7319438Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7320118Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7320260Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.7320339Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7320385Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7320486Z unimplemented []
2025-12-04T09:45:15.7320552Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7320660Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7321292Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7321332Z graph_break []
2025-12-04T09:45:15.7321411Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7321455Z Autotune Choices Stats:
2025-12-04T09:45:15.7322277Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.7322417Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7322540Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7322713Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7323401Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7324064Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7324720Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7325373Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7326028Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7326693Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7327350Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7328040Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7328714Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7329373Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7329514Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.7329556Z Autotune Choices Stats:
2025-12-04T09:45:15.7330381Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.7330651Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7330828Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7331149Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7331837Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7332545Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7333252Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7333936Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7334621Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7335302Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7335984Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7336663Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7337353Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7338045Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7338189Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.7338270Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7338315Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7338359Z unimplemented []
2025-12-04T09:45:15.7338423Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7338533Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7339160Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7339204Z graph_break []
2025-12-04T09:45:15.7339282Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7339328Z Autotune Choices Stats:
2025-12-04T09:45:15.7340135Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.7340274Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7340451Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7340625Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7341293Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7341979Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7342646Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7343303Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7343970Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7344627Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7345328Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7345989Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7346662Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7347340Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7347484Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.7347529Z Autotune Choices Stats:
2025-12-04T09:45:15.7348360Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.7348602Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7348782Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7349087Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7349781Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7350489Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7351184Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7351892Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7352581Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7353268Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7353958Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7354685Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7355366Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7356066Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7356226Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.7356305Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7356352Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7356393Z unimplemented []
2025-12-04T09:45:15.7356461Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7356570Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7357196Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7357235Z graph_break []
2025-12-04T09:45:15.7357318Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7357361Z Autotune Choices Stats:
2025-12-04T09:45:15.7358177Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.7358318Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7358440Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7358616Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7359314Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7359972Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7360685Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7361357Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7362014Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7362671Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7363334Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7364027Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7364684Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7365368Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7365520Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.7365563Z Autotune Choices Stats:
2025-12-04T09:45:15.7366391Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.7366633Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7366811Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7367118Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7367808Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7368505Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7369189Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7369886Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7370619Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7371301Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7371981Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7372686Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7373404Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7374079Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7374248Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.7374332Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7374377Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7374433Z unimplemented []
2025-12-04T09:45:15.7374499Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7374611Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7375232Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7375276Z graph_break []
2025-12-04T09:45:15.7375356Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7375400Z Autotune Choices Stats:
2025-12-04T09:45:15.7376218Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.7376360Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7376487Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7376666Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7377338Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7378034Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7378701Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7379376Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7380047Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7380776Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7381436Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7382092Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7382767Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7383423Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7383589Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.7383634Z Autotune Choices Stats:
2025-12-04T09:45:15.7384460Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.7384714Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7384896Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7385199Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7385890Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7386588Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7387283Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7387965Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7388674Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7389366Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7390060Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7390770Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7391453Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7392178Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7392320Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.7392402Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7392447Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7392488Z unimplemented []
2025-12-04T09:45:15.7392554Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7392660Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7393312Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7393367Z graph_break []
2025-12-04T09:45:15.7393449Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7393493Z Autotune Choices Stats:
2025-12-04T09:45:15.7394302Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.7394445Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7394569Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7394745Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7395414Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7396066Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7396762Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7397438Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7398101Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7398768Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7399434Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7400095Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7400775Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7401451Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7401592Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.7401634Z Autotune Choices Stats:
2025-12-04T09:45:15.7402473Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.7402737Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7402916Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7403221Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7403901Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7404580Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7405257Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7405949Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7406641Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7407331Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7408019Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7408699Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7409384Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7410062Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7410214Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.7410292Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7410339Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7410379Z unimplemented []
2025-12-04T09:45:15.7410478Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7410587Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7411219Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7411290Z graph_break []
2025-12-04T09:45:15.7411369Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7411415Z Autotune Choices Stats:
2025-12-04T09:45:15.7412221Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.7412378Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7412506Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7412678Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7413364Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7414020Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7414671Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7415341Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7416007Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7416671Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7417336Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7417987Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7418641Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7419296Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7419447Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:15.7419494Z Autotune Choices Stats:
2025-12-04T09:45:15.7420314Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.7420606Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7420800Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7421115Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7421802Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7422479Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7423160Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7423853Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7424565Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7425259Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7425942Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7426633Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7427331Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7428006Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7428144Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:15.7428225Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7428270Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7428312Z unimplemented []
2025-12-04T09:45:15.7428376Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7428487Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7429123Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7429163Z graph_break []
2025-12-04T09:45:15.7429244Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7429286Z Autotune Choices Stats:
2025-12-04T09:45:15.7430100Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.7430267Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7430390Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7430578Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7431245Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7431904Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7432562Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7433225Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7433880Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7434551Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7435224Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7435900Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7436557Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7437229Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7437372Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:15.7437414Z Autotune Choices Stats:
2025-12-04T09:45:15.7438253Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:15.7438492Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7438670Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7438983Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7439677Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7440370Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7441091Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7441775Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7442458Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7443173Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7443867Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7444564Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7445274Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7445957Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7446101Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:15.7446180Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7446226Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7446268Z unimplemented []
2025-12-04T09:45:15.7446335Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7446442Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7447066Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7447109Z graph_break []
2025-12-04T09:45:15.7447199Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7447244Z Autotune Choices Stats:
2025-12-04T09:45:15.7448048Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:15.7448200Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7448344Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7448518Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7449200Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7449874Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7450561Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7451222Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7451896Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7452549Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7453247Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7453934Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7454596Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7455271Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7455415Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:15.7455461Z Autotune Choices Stats:
2025-12-04T09:45:15.7456292Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:15.7456532Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7456725Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7457027Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7457732Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7458422Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7459108Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7459791Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7460525Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7461225Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7461904Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7462600Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7463307Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7463988Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7464132Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:15.7464214Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7464258Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7464303Z unimplemented []
2025-12-04T09:45:15.7464367Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7464478Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7465112Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7465155Z graph_break []
2025-12-04T09:45:15.7465236Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7465279Z Autotune Choices Stats:
2025-12-04T09:45:15.7466108Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:15.7466247Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7466373Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7466548Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7467239Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7467911Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7468584Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7469241Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7469896Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7470614Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7471274Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7471944Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7472629Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7473281Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7473427Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:15.7473470Z Autotune Choices Stats:
2025-12-04T09:45:15.7474290Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.7474528Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7474710Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7475023Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7475710Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7476395Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7477096Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7477795Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7478481Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7479163Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7479869Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7480589Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7481302Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7482003Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7482149Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:15.7482230Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7482280Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7482321Z unimplemented []
2025-12-04T09:45:15.7482391Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7482500Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7483121Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7483166Z graph_break []
2025-12-04T09:45:15.7483247Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7483294Z Autotune Choices Stats:
2025-12-04T09:45:15.7484096Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:15.7484238Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7484387Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7484564Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7485234Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7485909Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7486582Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7487261Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7487925Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7488592Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7489277Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7489939Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7490638Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7491320Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7491466Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:15.7491514Z Autotune Choices Stats:
2025-12-04T09:45:15.7492338Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:15.7492582Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7492764Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7493074Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7493777Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7494459Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7495153Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7495849Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7496536Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7497232Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7497908Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7498599Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7499285Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7499983Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7500156Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:15.7500241Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7500288Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7500332Z unimplemented []
2025-12-04T09:45:15.7500399Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7500555Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7501182Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7501227Z graph_break []
2025-12-04T09:45:15.7501310Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7501354Z Autotune Choices Stats:
2025-12-04T09:45:15.7502167Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:15.7502308Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7502436Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7502608Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7503293Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7503959Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7504644Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7505315Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7505978Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7506659Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7507326Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7507992Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7508661Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7509338Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7521881Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:15.7521928Z Autotune Choices Stats:
2025-12-04T09:45:15.7522757Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:15.7523003Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7523184Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7523489Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7524174Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7524882Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7525559Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7526265Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7527064Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7527754Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7528432Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7529116Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7529810Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7530540Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7530717Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:15.7530803Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7530849Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7530894Z unimplemented []
2025-12-04T09:45:15.7530974Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7531087Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7531713Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7531758Z graph_break []
2025-12-04T09:45:15.7531842Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7531885Z Autotune Choices Stats:
2025-12-04T09:45:15.7532703Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:15.7532844Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7532977Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7533155Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7533840Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7534537Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7535207Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7535892Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7536573Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7537254Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7537930Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7538600Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7539284Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7539952Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7540119Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:15.7540168Z Autotune Choices Stats:
2025-12-04T09:45:15.7541046Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:15.7541311Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7541499Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7541811Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7542505Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7543194Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7543917Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7544609Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7545340Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7546047Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7546753Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7547452Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7548137Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7548844Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7548992Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:15.7549097Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.7549152Z Traceback (most recent call last):
2025-12-04T09:45:15.7549324Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.7549372Z     self.assertTrue(
2025-12-04T09:45:15.7549499Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.7549570Z     raise self.failureException(msg)
2025-12-04T09:45:15.7549710Z AssertionError: False is not true : Log file /tmp/tmpkqvme528/flex_attention_configs.json was not created
2025-12-04T09:45:15.7549714Z 
2025-12-04T09:45:15.7549802Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.7549998Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.7550001Z 
2025-12-04T09:45:15.7550103Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.7550185Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7550236Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7550277Z unimplemented []
2025-12-04T09:45:15.7550348Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7551044Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.7551155Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7551199Z graph_break []
2025-12-04T09:45:15.7551280Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7551825Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.7551879Z   current_size = base.storage().size()
2025-12-04T09:45:15.7551927Z Autotune Choices Stats:
2025-12-04T09:45:15.7552742Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.7552886Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7553036Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7553211Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7553879Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7554575Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7555240Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7555899Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7556551Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7557202Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7557872Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7558519Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7559183Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7559862Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7560006Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.7560053Z Autotune Choices Stats:
2025-12-04T09:45:15.7560922Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.7561160Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7561342Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7561644Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7562353Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7563028Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7563729Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7564417Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7565094Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7565747Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7566380Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7567022Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7567655Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7568296Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7568449Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.7568524Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7568569Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7568607Z unimplemented []
2025-12-04T09:45:15.7568669Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7568770Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7569357Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7569395Z graph_break []
2025-12-04T09:45:15.7569469Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7569509Z Autotune Choices Stats:
2025-12-04T09:45:15.7570269Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.7570403Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7570551Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7570715Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7571348Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7572030Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7572680Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7573295Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7573904Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7574516Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7575128Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7575746Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7576353Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7576984Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7577133Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.7577173Z Autotune Choices Stats:
2025-12-04T09:45:15.7577930Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.7578153Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7578320Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7578603Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7579241Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7579876Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7580532Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7581172Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7581826Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7582457Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7583086Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7583716Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7584353Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7584980Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7585121Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.7585207Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7585250Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7585290Z unimplemented []
2025-12-04T09:45:15.7585351Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7585466Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7586045Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7586088Z graph_break []
2025-12-04T09:45:15.7586164Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7586207Z Autotune Choices Stats:
2025-12-04T09:45:15.7586966Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.7587098Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7587220Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7587381Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7587991Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7588609Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7589214Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7589842Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7590487Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7591103Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7591708Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7592321Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7592962Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7593570Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7593726Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.7593771Z Autotune Choices Stats:
2025-12-04T09:45:15.7594533Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.7594767Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7594937Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7595216Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7595854Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7596477Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7597121Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7597748Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7598388Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7599046Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7599680Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7600309Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7600969Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7601611Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7601741Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.7601818Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7601861Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7601899Z unimplemented []
2025-12-04T09:45:15.7601960Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7602060Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7602673Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7602727Z graph_break []
2025-12-04T09:45:15.7602803Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7602843Z Autotune Choices Stats:
2025-12-04T09:45:15.7603604Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.7603737Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7603852Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7604017Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7604638Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7605255Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7605878Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7606498Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7607115Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7607732Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7608341Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7608952Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7609562Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7610183Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7610317Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.7610357Z Autotune Choices Stats:
2025-12-04T09:45:15.7611164Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.7611409Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7611575Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7611858Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7612499Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7613138Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7613770Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7614414Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7615044Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7618693Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7619340Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7619971Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7620647Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7621279Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7621412Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.7621514Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7621560Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7621598Z unimplemented []
2025-12-04T09:45:15.7621660Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7621764Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7622342Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7622396Z graph_break []
2025-12-04T09:45:15.7622490Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7622532Z Autotune Choices Stats:
2025-12-04T09:45:15.7623288Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.7623432Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7623554Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7623717Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7624332Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7624936Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7625542Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7626160Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7626777Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7627399Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7628032Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7628648Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7629269Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7629900Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7630032Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.7630083Z Autotune Choices Stats:
2025-12-04T09:45:15.7630882Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.7631103Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7631298Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7631589Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7632230Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7632858Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7633482Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7634110Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7634769Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7635411Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7636047Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7636693Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7637334Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7637973Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7638104Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.7638181Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7638222Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7638262Z unimplemented []
2025-12-04T09:45:15.7638325Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7638426Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7639032Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7639069Z graph_break []
2025-12-04T09:45:15.7639145Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7639184Z Autotune Choices Stats:
2025-12-04T09:45:15.7639963Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.7640114Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7640232Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7640395Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7641035Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7641648Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7642264Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7642871Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7643496Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7644124Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7644752Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7645379Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7645996Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7646613Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7646745Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.7646784Z Autotune Choices Stats:
2025-12-04T09:45:15.7647572Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.7647800Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7647968Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7648265Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7648920Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7649570Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7650205Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7650878Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7651518Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7652175Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7652823Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7653473Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7654117Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7654757Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7654888Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.7654964Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7655007Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7655043Z unimplemented []
2025-12-04T09:45:15.7655105Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7655205Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7655796Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7655835Z graph_break []
2025-12-04T09:45:15.7655910Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7655976Z Autotune Choices Stats:
2025-12-04T09:45:15.7656730Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.7656861Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7656998Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7657164Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7657803Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7658418Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7659032Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7659649Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7660278Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7660914Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7661544Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7662176Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7662820Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7663433Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7663562Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.7663604Z Autotune Choices Stats:
2025-12-04T09:45:15.7664376Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.7664602Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7664783Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7665069Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7665727Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7666371Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7667013Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7667650Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7668294Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7668931Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7669567Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7670222Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7670897Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7671548Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7671677Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.7671751Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7671792Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7671831Z unimplemented []
2025-12-04T09:45:15.7671890Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7671991Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7672574Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7672611Z graph_break []
2025-12-04T09:45:15.7672684Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7672723Z Autotune Choices Stats:
2025-12-04T09:45:15.7673492Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.7673622Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7673738Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7673901Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7674552Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7675184Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7675803Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7676422Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7677041Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7677664Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7678285Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7678909Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7679543Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7680159Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7680291Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.7680331Z Autotune Choices Stats:
2025-12-04T09:45:15.7681127Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.7681350Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7681517Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7681803Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7682468Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7683110Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7683760Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7684407Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7685044Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7685685Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7686337Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7686975Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7687617Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7688278Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7688410Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.7688486Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7688527Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7688564Z unimplemented []
2025-12-04T09:45:15.7688627Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7688729Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7689306Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7689342Z graph_break []
2025-12-04T09:45:15.7689416Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7689457Z Autotune Choices Stats:
2025-12-04T09:45:15.7690290Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.7690454Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7690586Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7690747Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7691360Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7691980Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7692614Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7693237Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7693859Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7694466Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7695087Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7695693Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7696315Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7696937Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7697066Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.7697107Z Autotune Choices Stats:
2025-12-04T09:45:15.7697872Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.7698091Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7698260Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7698540Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7699186Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7699810Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7700482Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7701133Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7701766Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7702396Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7703027Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7703671Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7704298Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7704942Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7705088Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.7705163Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7705204Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7705241Z unimplemented []
2025-12-04T09:45:15.7705302Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7705402Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7705977Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7706015Z graph_break []
2025-12-04T09:45:15.7706088Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7706130Z Autotune Choices Stats:
2025-12-04T09:45:15.7706881Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.7707009Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7707124Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7707284Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7707913Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7708517Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7709131Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7709755Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7710356Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7711008Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7711614Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7712237Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7712847Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7713457Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7713609Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.7713649Z Autotune Choices Stats:
2025-12-04T09:45:15.7714413Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.7714633Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7714801Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7715077Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7715712Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7716347Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7716968Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7717600Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7718245Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7718880Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7719516Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7720138Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7720825Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7721453Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7721584Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.7721684Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7721727Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7721764Z unimplemented []
2025-12-04T09:45:15.7721825Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7721938Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7722515Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7722552Z graph_break []
2025-12-04T09:45:15.7722627Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7722667Z Autotune Choices Stats:
2025-12-04T09:45:15.7723412Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.7723541Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7723657Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7723817Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7724424Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7725046Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7725655Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7726280Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7726896Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7727501Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7728112Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7728717Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7729334Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7729938Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7730067Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.7730129Z Autotune Choices Stats:
2025-12-04T09:45:15.7730944Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.7731175Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7731345Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7731623Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7732257Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7732900Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7733538Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7734157Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7734795Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7735453Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7736079Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7736710Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7737341Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7737979Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7738109Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.7738183Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7738224Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7738261Z unimplemented []
2025-12-04T09:45:15.7738320Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7738421Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7739013Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7739072Z graph_break []
2025-12-04T09:45:15.7739144Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7739184Z Autotune Choices Stats:
2025-12-04T09:45:15.7739929Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.7740057Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7740173Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7740333Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7740986Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7741587Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7742215Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7742824Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7743460Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7744077Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7744692Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7745316Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7745937Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7746556Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7746687Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.7746728Z Autotune Choices Stats:
2025-12-04T09:45:15.7747515Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.7747743Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7747918Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7748196Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7748829Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7749461Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7750085Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7750761Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7751393Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7752053Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7752690Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7753320Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7753949Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7754568Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7754698Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.7754774Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7754828Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7754866Z unimplemented []
2025-12-04T09:45:15.7754927Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7755026Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7755603Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7755639Z graph_break []
2025-12-04T09:45:15.7755724Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7755774Z Autotune Choices Stats:
2025-12-04T09:45:15.7756509Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.7756649Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7756766Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7756927Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7757548Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7758157Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7758765Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7759382Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7759995Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7760643Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7761273Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7761881Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7762489Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7763098Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7763227Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:15.7763267Z Autotune Choices Stats:
2025-12-04T09:45:15.7764040Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.7764262Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7764452Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7764731Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7765379Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7766008Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7766631Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7767265Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7767909Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7768537Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7769184Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7769822Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7770484Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7771110Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7771239Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:15.7771317Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7771358Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7771396Z unimplemented []
2025-12-04T09:45:15.7771456Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7771556Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7772150Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7772189Z graph_break []
2025-12-04T09:45:15.7772262Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7772304Z Autotune Choices Stats:
2025-12-04T09:45:15.7773062Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.7773216Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7773332Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7773493Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7774105Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7774712Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7775315Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7775920Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7776536Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7777159Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7777778Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7778395Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7779003Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7779617Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7779747Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:15.7779791Z Autotune Choices Stats:
2025-12-04T09:45:15.7780583Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:15.7780804Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7780974Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7781253Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7781916Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7782558Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7783185Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7783811Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7784443Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7785097Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7785733Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7786376Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7787028Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7787657Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7787789Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:15.7787869Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7787913Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7787952Z unimplemented []
2025-12-04T09:45:15.7788015Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7788117Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7788698Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7788735Z graph_break []
2025-12-04T09:45:15.7788814Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7788869Z Autotune Choices Stats:
2025-12-04T09:45:15.7789615Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:15.7789747Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7789889Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7790054Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7790707Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7791336Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7791948Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7792558Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7793165Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7793809Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7794434Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7795055Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7795678Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7796288Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7796421Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:15.7796463Z Autotune Choices Stats:
2025-12-04T09:45:15.7797237Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:15.7797460Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7797642Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7797925Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7798567Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7799212Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7799864Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7800515Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7801145Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7801783Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7802431Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7803084Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7803728Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7804371Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7804504Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:15.7804578Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7804622Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7804660Z unimplemented []
2025-12-04T09:45:15.7804723Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7804824Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7805406Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7805447Z graph_break []
2025-12-04T09:45:15.7805521Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7805565Z Autotune Choices Stats:
2025-12-04T09:45:15.7806323Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:15.7806457Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7806575Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7806738Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7807373Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7808008Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7808621Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7809243Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7809866Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7810551Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7811158Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7811791Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7812416Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7813042Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7813174Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:15.7813219Z Autotune Choices Stats:
2025-12-04T09:45:15.7813985Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.7814208Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7814379Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7814659Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7815310Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7815955Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7816600Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7817240Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7817879Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7818523Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7819153Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7819811Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7820498Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7821142Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7821291Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:15.7821371Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7821415Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7821458Z unimplemented []
2025-12-04T09:45:15.7821520Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7821623Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7822212Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7822252Z graph_break []
2025-12-04T09:45:15.7822331Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7822373Z Autotune Choices Stats:
2025-12-04T09:45:15.7823132Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:15.7823263Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7823380Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7823566Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7824186Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7824810Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7825449Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7826058Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7826665Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7827548Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7828191Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7828803Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7829427Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7830052Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7830186Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:15.7830228Z Autotune Choices Stats:
2025-12-04T09:45:15.7831037Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:15.7831258Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7831430Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7831720Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7832353Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7833008Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7833665Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7834319Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7834967Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7835600Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7836229Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7836884Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7837523Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7838172Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7838327Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:15.7838403Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7838449Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7838487Z unimplemented []
2025-12-04T09:45:15.7838552Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7838655Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7839239Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7839283Z graph_break []
2025-12-04T09:45:15.7839358Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7839402Z Autotune Choices Stats:
2025-12-04T09:45:15.7840148Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:15.7840281Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7840400Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7840596Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7841229Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7841839Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7842465Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7843109Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7843712Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7844318Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7844931Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7845549Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7846157Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7846774Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7846928Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:15.7846975Z Autotune Choices Stats:
2025-12-04T09:45:15.7847732Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:15.7847954Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7848127Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7848410Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7849053Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7849712Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7850337Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7851025Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7851688Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7852328Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7852961Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7853597Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7854246Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7854875Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7855004Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:15.7855094Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7855147Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7855189Z unimplemented []
2025-12-04T09:45:15.7855249Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7855354Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7855940Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7855980Z graph_break []
2025-12-04T09:45:15.7856056Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7856098Z Autotune Choices Stats:
2025-12-04T09:45:15.7856838Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:15.7856966Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7857086Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7857252Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7857867Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7858492Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7859102Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7859729Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7860352Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7861009Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7861632Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7862246Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7862871Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7863480Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7863613Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:15.7863668Z Autotune Choices Stats:
2025-12-04T09:45:15.7864448Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:15.7864686Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7864857Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7865140Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7865771Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7866397Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7867040Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7867668Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7868311Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7868957Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7869587Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7870225Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7870881Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7871539Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7871672Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:15.7871747Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7871793Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7871831Z unimplemented []
2025-12-04T09:45:15.7871894Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7871996Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7872593Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7872660Z graph_break []
2025-12-04T09:45:15.7872733Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7872791Z Autotune Choices Stats:
2025-12-04T09:45:15.7873536Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:15.7873670Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7873786Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7873951Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7874570Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7875179Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7875814Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7876422Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7877054Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7877673Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7878290Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7878909Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7879529Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7880157Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7880308Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:15.7880473Z Autotune Choices Stats:
2025-12-04T09:45:15.7881248Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.7881491Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7881672Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7881952Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7882588Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7883219Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7883858Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7884507Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7885146Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7885803Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7886457Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7887095Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7887729Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7888367Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7888496Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:15.7888593Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.7888653Z Traceback (most recent call last):
2025-12-04T09:45:15.7888814Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.7888855Z     self.assertTrue(
2025-12-04T09:45:15.7888968Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.7889018Z     raise self.failureException(msg)
2025-12-04T09:45:15.7889148Z AssertionError: False is not true : Log file /tmp/tmp7xa3g518/flex_attention_configs.json was not created
2025-12-04T09:45:15.7889154Z 
2025-12-04T09:45:15.7889230Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.7889404Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.7889406Z 
2025-12-04T09:45:15.7889501Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.7889600Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7889647Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7889686Z unimplemented []
2025-12-04T09:45:15.7889751Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7890348Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.7890491Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7890529Z graph_break []
2025-12-04T09:45:15.7890608Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7891108Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.7891159Z   current_size = base.storage().size()
2025-12-04T09:45:15.7891199Z Autotune Choices Stats:
2025-12-04T09:45:15.7891961Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.7892094Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7892209Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7892373Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7893008Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7893619Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7894236Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7894868Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7895474Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7896079Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7896689Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7897308Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7897914Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7898549Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7898706Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.7898746Z Autotune Choices Stats:
2025-12-04T09:45:15.7899509Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.7899734Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7899903Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7900184Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7900855Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7901495Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7902124Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7902768Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7903420Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7904050Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7904678Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7905313Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7905953Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7906579Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7906711Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.7906801Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7906854Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7906895Z unimplemented []
2025-12-04T09:45:15.7906956Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7907060Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7907649Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7907689Z graph_break []
2025-12-04T09:45:15.7907763Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7907811Z Autotune Choices Stats:
2025-12-04T09:45:15.7908558Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.7908687Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7908805Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7908966Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7909579Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7910203Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7910845Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7911474Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7912106Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7912723Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7913333Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7913941Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7914574Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7915175Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7915305Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.7915364Z Autotune Choices Stats:
2025-12-04T09:45:15.7916148Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.7916382Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7916553Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7916834Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7917471Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7918104Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7918751Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7919376Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7920012Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7920686Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7921312Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7921943Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7922572Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7923215Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7923349Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.7923423Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7923471Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7923508Z unimplemented []
2025-12-04T09:45:15.7923571Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7923672Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7924266Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7924332Z graph_break []
2025-12-04T09:45:15.7924409Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7924450Z Autotune Choices Stats:
2025-12-04T09:45:15.7925197Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.7925328Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7925444Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7925607Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7926226Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7926839Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7927459Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7928068Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7928695Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7929328Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7929936Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7930575Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7931186Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7931819Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7931953Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.7931996Z Autotune Choices Stats:
2025-12-04T09:45:15.7932777Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.7933012Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7933192Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7933474Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7934120Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7934756Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7935394Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7936046Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7936690Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7937341Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7938019Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7938661Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7939306Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7939978Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7940112Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.7940189Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7940235Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7940289Z unimplemented []
2025-12-04T09:45:15.7940350Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7940493Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7941080Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7941120Z graph_break []
2025-12-04T09:45:15.7941195Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7941254Z Autotune Choices Stats:
2025-12-04T09:45:15.7942016Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.7942162Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7942281Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7942451Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7943071Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7943693Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7944312Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7944946Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7945562Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7946205Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7946834Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7947455Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7948073Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7948688Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7948819Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.7948864Z Autotune Choices Stats:
2025-12-04T09:45:15.7949664Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.7949892Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7950075Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7950372Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7951065Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7951705Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7952350Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7952993Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7953650Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7954288Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7954951Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7955604Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7956244Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7956881Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7957011Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.7957090Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7957136Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7957173Z unimplemented []
2025-12-04T09:45:15.7957237Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7957339Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7957935Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7957974Z graph_break []
2025-12-04T09:45:15.7958051Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7958092Z Autotune Choices Stats:
2025-12-04T09:45:15.7958862Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.7959006Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7959137Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7959302Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7959944Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7960579Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7961201Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7961816Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7962446Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7963059Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7963709Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7964340Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7964960Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7965584Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7965720Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.7965761Z Autotune Choices Stats:
2025-12-04T09:45:15.7966559Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.7966786Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7966954Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7967245Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7967908Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7968556Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7969196Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7969838Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7970502Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7971154Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7971797Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7972465Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7973115Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7973751Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7973884Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.7973958Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7974005Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7974045Z unimplemented []
2025-12-04T09:45:15.7974109Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7974210Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7974796Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.7974835Z graph_break []
2025-12-04T09:45:15.7974909Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7974954Z Autotune Choices Stats:
2025-12-04T09:45:15.7975730Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.7975865Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7975985Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7976170Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7976799Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7977428Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7978045Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7978656Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7979274Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7979911Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7980573Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7981203Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7981836Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7982452Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7982584Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.7982629Z Autotune Choices Stats:
2025-12-04T09:45:15.7983407Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.7983635Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7983818Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7984105Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7984753Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7985409Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7986055Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7986691Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7987333Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7987977Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7988626Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7989279Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7989922Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7990608Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7990739Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.7990819Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.7990862Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.7990903Z unimplemented []
2025-12-04T09:45:15.7990964Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.7991069Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.7991656Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.7991695Z graph_break []
2025-12-04T09:45:15.7991774Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.7991814Z Autotune Choices Stats:
2025-12-04T09:45:15.7992592Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.7992726Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.7992843Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.7993010Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.7993665Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7994297Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7994935Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7995568Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7996182Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7996799Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.7997434Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7998065Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7998688Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7999318Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.7999455Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.7999497Z Autotune Choices Stats:
2025-12-04T09:45:15.8000275Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.8000540Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8000710Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8000997Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8001656Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8002310Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8002959Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8003612Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8004251Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8004899Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8005531Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8006189Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8006841Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8007483Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8007627Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.8007701Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8007747Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8007785Z unimplemented []
2025-12-04T09:45:15.8007849Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8007951Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8008535Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8008576Z graph_break []
2025-12-04T09:45:15.8008649Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8008694Z Autotune Choices Stats:
2025-12-04T09:45:15.8009452Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.8009586Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8009704Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8009883Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8010535Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8011157Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8011813Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8012430Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8013046Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8013662Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8014297Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8014949Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8015572Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8018063Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8018210Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.8018254Z Autotune Choices Stats:
2025-12-04T09:45:15.8019027Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.8019253Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8019423Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8019709Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8020353Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8021039Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8021695Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8022344Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8022998Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8023639Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8024273Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8024921Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8025570Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8026212Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8026354Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.8026444Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8026488Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8026529Z unimplemented []
2025-12-04T09:45:15.8026590Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8026696Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8027292Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8027330Z graph_break []
2025-12-04T09:45:15.8027406Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8027447Z Autotune Choices Stats:
2025-12-04T09:45:15.8028213Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.8028343Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8028462Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8028628Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8029269Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8029888Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8030549Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8031187Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8031799Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8032413Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8033037Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8033666Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8034282Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8034910Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8035056Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.8035109Z Autotune Choices Stats:
2025-12-04T09:45:15.8035883Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.8036107Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8036277Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8036565Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8037201Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8037841Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8038496Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8039147Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8039798Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8040478Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8041116Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8041772Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8042426Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8043063Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8043196Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.8043270Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8043328Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8043379Z unimplemented []
2025-12-04T09:45:15.8043445Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8043546Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8044142Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8044182Z graph_break []
2025-12-04T09:45:15.8044256Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8044301Z Autotune Choices Stats:
2025-12-04T09:45:15.8045061Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.8045194Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8045310Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8045481Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8046106Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8046735Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8047358Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8047984Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8048622Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8049238Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8049861Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8050518Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8051156Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8051767Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8051903Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.8051947Z Autotune Choices Stats:
2025-12-04T09:45:15.8052747Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.8052988Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8053158Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8053447Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8054090Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8054726Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8055375Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8056008Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8056658Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8057319Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8057954Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8058593Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8059236Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8059885Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8060016Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.8060095Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8060137Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8060177Z unimplemented []
2025-12-04T09:45:15.8060237Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8060341Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8060976Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8061052Z graph_break []
2025-12-04T09:45:15.8061128Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8061169Z Autotune Choices Stats:
2025-12-04T09:45:15.8061933Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.8062063Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8062184Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8062347Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8062974Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8063599Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8064232Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8064848Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8065473Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8066106Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8066726Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8067338Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8067956Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8068581Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8068713Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.8068757Z Autotune Choices Stats:
2025-12-04T09:45:15.8069551Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.8069786Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8069966Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8070252Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8070932Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8071571Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8072205Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8072860Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8073499Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8074147Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8074805Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8075449Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8076087Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8076724Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8076858Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.8076932Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8076978Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8077017Z unimplemented []
2025-12-04T09:45:15.8077090Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8077192Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8077777Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8077816Z graph_break []
2025-12-04T09:45:15.8077892Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8077943Z Autotune Choices Stats:
2025-12-04T09:45:15.8078720Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.8078862Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8078978Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8079149Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8079778Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8080390Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8081023Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8081657Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8082270Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8082904Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8083537Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8084157Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8084784Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8085391Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8085527Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.8085568Z Autotune Choices Stats:
2025-12-04T09:45:15.8086349Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.8086576Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8086754Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8087049Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8087708Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8088350Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8088984Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8089626Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8090278Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8090949Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8091598Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8092263Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8092908Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8093563Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8093695Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.8093774Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8093816Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8093856Z unimplemented []
2025-12-04T09:45:15.8093917Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8094021Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8094620Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8094661Z graph_break []
2025-12-04T09:45:15.8094736Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8094779Z Autotune Choices Stats:
2025-12-04T09:45:15.8095555Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.8095696Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8095826Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8095992Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8096632Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8097249Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8097885Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8098523Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8099151Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8099764Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8100442Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8101073Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8101685Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8102316Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8102447Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:15.8102491Z Autotune Choices Stats:
2025-12-04T09:45:15.8103277Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.8103501Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8103672Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8103959Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8104617Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8105279Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8105918Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8106556Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8107206Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8107857Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8108494Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8109155Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8109808Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8110488Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8110621Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:15.8110697Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8110742Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8110781Z unimplemented []
2025-12-04T09:45:15.8110847Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8110947Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8111532Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8111571Z graph_break []
2025-12-04T09:45:15.8111648Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8111690Z Autotune Choices Stats:
2025-12-04T09:45:15.8112470Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.8112605Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8112720Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8112912Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8113533Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8114169Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8114791Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8115408Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8116022Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8116660Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8117280Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8117918Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8118546Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8119164Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8119297Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:15.8119337Z Autotune Choices Stats:
2025-12-04T09:45:15.8120115Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:15.8120345Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8120557Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8120842Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8121490Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8122150Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8122801Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8123438Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8124079Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8124719Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8125361Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8126000Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8126661Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8127313Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8127446Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:15.8127528Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8127571Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8127611Z unimplemented []
2025-12-04T09:45:15.8127671Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8127775Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8128363Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8128404Z graph_break []
2025-12-04T09:45:15.8128479Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8128523Z Autotune Choices Stats:
2025-12-04T09:45:15.8129287Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:15.8129417Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8133947Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8134114Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8134759Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8135384Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8136005Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8136613Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8137220Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8137824Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8138449Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8139072Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8139689Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8140302Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8140470Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:15.8140512Z Autotune Choices Stats:
2025-12-04T09:45:15.8141268Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:15.8141490Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8141660Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8141945Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8142593Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8143222Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8143873Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8144509Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8145136Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8145767Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8146395Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8147041Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8147684Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8148330Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8148470Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:15.8148545Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8148588Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8148625Z unimplemented []
2025-12-04T09:45:15.8148687Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8148787Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8149370Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8149407Z graph_break []
2025-12-04T09:45:15.8149480Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8149521Z Autotune Choices Stats:
2025-12-04T09:45:15.8150261Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:15.8150390Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8150540Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8150716Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8151332Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8151952Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8152574Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8153194Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8153800Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8154405Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8155015Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8155631Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8156247Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8156859Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8156999Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:15.8157040Z Autotune Choices Stats:
2025-12-04T09:45:15.8157809Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.8158029Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8158199Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8158481Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8159116Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8159792Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8160467Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8161104Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8161746Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8162376Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8162995Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8163626Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8164267Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8164905Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8165042Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:15.8165129Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8165171Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8165208Z unimplemented []
2025-12-04T09:45:15.8165269Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8165369Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8165950Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8165988Z graph_break []
2025-12-04T09:45:15.8166060Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8166101Z Autotune Choices Stats:
2025-12-04T09:45:15.8166850Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:15.8166978Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8167093Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8167254Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8167880Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8168485Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8169108Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8169729Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8172316Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8172933Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8173591Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8174203Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8174835Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8175443Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8175591Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:15.8175644Z Autotune Choices Stats:
2025-12-04T09:45:15.8176409Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:15.8176629Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8176853Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8177134Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8177777Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8178404Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8179041Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8179670Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8180309Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8180980Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8181641Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8182271Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8182910Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8183537Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8183667Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:15.8183742Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8183797Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8183835Z unimplemented []
2025-12-04T09:45:15.8183895Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8183997Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8184571Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8184620Z graph_break []
2025-12-04T09:45:15.8184693Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8184733Z Autotune Choices Stats:
2025-12-04T09:45:15.8185498Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:15.8185627Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8185742Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8185902Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8186517Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8187138Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8187747Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8188347Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8188970Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8189591Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8190199Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8190824Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8191446Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8192054Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8192183Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:15.8192223Z Autotune Choices Stats:
2025-12-04T09:45:15.8192998Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:15.8193242Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8193413Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8193693Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8194341Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8194971Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8195607Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8196233Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8196863Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8197502Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8198161Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8198789Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8199424Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8200065Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8200193Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:15.8200270Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8200311Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8200348Z unimplemented []
2025-12-04T09:45:15.8200432Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8200535Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8201107Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8201170Z graph_break []
2025-12-04T09:45:15.8201258Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8201299Z Autotune Choices Stats:
2025-12-04T09:45:15.8202045Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:15.8202176Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8202303Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8202464Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8203081Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8203693Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8204315Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8204924Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8205528Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8206155Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8206780Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8207384Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8207994Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8208613Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8208748Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:15.8208788Z Autotune Choices Stats:
2025-12-04T09:45:15.8209557Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:15.8209787Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8209964Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8210243Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8210933Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8211559Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8212185Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8212838Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8213469Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8214097Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8214757Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8215400Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8216028Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8216657Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8216785Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:15.8216858Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8216900Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8216937Z unimplemented []
2025-12-04T09:45:15.8216998Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8217111Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8217689Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8217726Z graph_break []
2025-12-04T09:45:15.8217798Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8217838Z Autotune Choices Stats:
2025-12-04T09:45:15.8218594Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:15.8218734Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8218849Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8219010Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8219640Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8220247Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8220894Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8221511Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8222116Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8222721Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8223360Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8223982Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8224590Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8225196Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8225325Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:15.8225366Z Autotune Choices Stats:
2025-12-04T09:45:15.8226144Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.8226363Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8226531Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8226818Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8227465Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8228104Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8228726Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8229351Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8229995Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8230665Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8231286Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8231951Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8232614Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8233242Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8233369Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:15.8233446Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8233488Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8233525Z unimplemented []
2025-12-04T09:45:15.8233585Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8233685Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8234282Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8234319Z graph_break []
2025-12-04T09:45:15.8234392Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8234433Z Autotune Choices Stats:
2025-12-04T09:45:15.8235178Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:15.8235317Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8235435Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8235604Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8236219Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8236834Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8237446Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8238052Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8238669Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8239277Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8239902Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8240545Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8241168Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8241776Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8241906Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:15.8241946Z Autotune Choices Stats:
2025-12-04T09:45:15.8242712Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.8242943Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8243112Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8243392Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8244030Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8244680Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8245327Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8245958Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8246587Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8247225Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8247851Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8248480Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8249128Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8249762Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8249892Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:15.8249985Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.8250032Z Traceback (most recent call last):
2025-12-04T09:45:15.8250188Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.8250229Z     self.assertTrue(
2025-12-04T09:45:15.8250334Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.8250383Z     raise self.failureException(msg)
2025-12-04T09:45:15.8250551Z AssertionError: False is not true : Log file /tmp/tmp5m0xxtnv/flex_attention_configs.json was not created
2025-12-04T09:45:15.8250555Z 
2025-12-04T09:45:15.8250632Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.8250798Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.8250801Z 
2025-12-04T09:45:15.8250888Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.8250963Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8251005Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8251043Z unimplemented []
2025-12-04T09:45:15.8251117Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8251702Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.8251801Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8251837Z graph_break []
2025-12-04T09:45:15.8251909Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8252425Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.8252490Z   current_size = base.storage().size()
2025-12-04T09:45:15.8252529Z Autotune Choices Stats:
2025-12-04T09:45:15.8253295Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.8253425Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8253554Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8253714Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8254324Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8254930Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8255550Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8256155Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8256762Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8257384Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8258004Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8258609Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8259214Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8259835Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8259965Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.8260005Z Autotune Choices Stats:
2025-12-04T09:45:15.8260806Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.8261047Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8261232Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8261518Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8262162Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8262790Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8263417Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8264060Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8264689Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8265317Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8265966Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8266605Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8267234Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8267864Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8267995Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.8268069Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8268111Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8268151Z unimplemented []
2025-12-04T09:45:15.8268211Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8268321Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8268900Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8268940Z graph_break []
2025-12-04T09:45:15.8269012Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8269054Z Autotune Choices Stats:
2025-12-04T09:45:15.8269806Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.8269945Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8270061Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8270222Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8270985Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8271593Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8272194Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8272808Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8273414Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8274022Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8274655Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8275275Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8275885Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8276508Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8276638Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.8276679Z Autotune Choices Stats:
2025-12-04T09:45:15.8277458Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.8277682Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8277852Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8278147Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8278793Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8279432Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8280061Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8280752Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8281401Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8282035Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8282661Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8283315Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8283955Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8284588Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8284718Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.8284793Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8284836Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8284875Z unimplemented []
2025-12-04T09:45:15.8284934Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8285037Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8285620Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8285657Z graph_break []
2025-12-04T09:45:15.8285731Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8285772Z Autotune Choices Stats:
2025-12-04T09:45:15.8286522Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.8286663Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8286777Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8286951Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8287564Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8288189Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8288795Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8289407Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8290030Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8290661Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8291284Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8291899Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8292521Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8293125Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8293255Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.8293296Z Autotune Choices Stats:
2025-12-04T09:45:15.8294068Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.8294312Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8294479Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8294762Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8295407Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8296062Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8296704Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8297331Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8297963Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8298609Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8299245Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8299876Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8300573Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8301215Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8301347Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.8301420Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8301464Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8301503Z unimplemented []
2025-12-04T09:45:15.8301565Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8301667Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8302248Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8302286Z graph_break []
2025-12-04T09:45:15.8302359Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8302400Z Autotune Choices Stats:
2025-12-04T09:45:15.8303167Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.8303297Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8303413Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8303575Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8304209Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8304827Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8305453Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8306063Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8306679Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8307298Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8307908Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8308532Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8309151Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8309771Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8309902Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.8309943Z Autotune Choices Stats:
2025-12-04T09:45:15.8310761Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.8310982Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8311150Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8311444Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8312085Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8312716Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8313364Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8314009Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8314645Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8315277Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8315911Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8316547Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8317197Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8317832Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8317961Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.8318038Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8318091Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8318131Z unimplemented []
2025-12-04T09:45:15.8318192Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8318294Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8318871Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8318910Z graph_break []
2025-12-04T09:45:15.8318985Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8319026Z Autotune Choices Stats:
2025-12-04T09:45:15.8319776Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.8319914Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8320031Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8320197Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8320855Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8321480Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8322103Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8322726Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8323328Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8323936Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8324553Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8325156Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8325775Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8326398Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8326535Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.8326575Z Autotune Choices Stats:
2025-12-04T09:45:15.8327357Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.8327578Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8327747Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8328027Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8328676Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8329300Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8329937Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8330610Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8331260Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8331891Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8332519Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8333169Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8333797Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8334437Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8334580Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.8334654Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8334699Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8334738Z unimplemented []
2025-12-04T09:45:15.8334800Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8334901Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8335488Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8335529Z graph_break []
2025-12-04T09:45:15.8335602Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8335645Z Autotune Choices Stats:
2025-12-04T09:45:15.8336388Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.8336520Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8336637Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8336803Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8337445Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8338049Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8338668Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8339288Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8339905Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8340549Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8341169Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8341797Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8342405Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8343041Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8343185Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.8343228Z Autotune Choices Stats:
2025-12-04T09:45:15.8344005Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.8344230Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8344396Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8344680Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8345322Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8345963Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8346589Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8347229Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8347871Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8348510Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8349131Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8349764Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8350455Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8351088Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8351228Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.8351304Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8351359Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8351400Z unimplemented []
2025-12-04T09:45:15.8351461Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8351564Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8352142Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8352181Z graph_break []
2025-12-04T09:45:15.8352256Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8352297Z Autotune Choices Stats:
2025-12-04T09:45:15.8353062Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.8353192Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8353308Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8353471Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8354083Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8354699Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8355311Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8355929Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8356545Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8357167Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8357782Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8358388Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8359012Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8359616Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8359754Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.8359795Z Autotune Choices Stats:
2025-12-04T09:45:15.8360598Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.8360818Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8361004Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8361289Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8361923Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8362568Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8363207Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8363838Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8364478Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8365124Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8365765Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8366400Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8367030Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8367664Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8367794Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.8367868Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8367912Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8367949Z unimplemented []
2025-12-04T09:45:15.8368024Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8368125Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8368706Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8368757Z graph_break []
2025-12-04T09:45:15.8368831Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8368873Z Autotune Choices Stats:
2025-12-04T09:45:15.8369634Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.8369766Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8369879Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8370043Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8370704Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8371328Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8371939Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8372547Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8373165Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8373789Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8374414Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8375021Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8375630Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8376249Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8376380Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.8376419Z Autotune Choices Stats:
2025-12-04T09:45:15.8377182Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.8377421Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8377589Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8377870Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8378515Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8379140Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8379771Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8380436Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8381065Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8381704Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8382344Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8382988Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8383621Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8384267Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8384397Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.8384472Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8384514Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8384554Z unimplemented []
2025-12-04T09:45:15.8384614Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8384716Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8385295Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8385348Z graph_break []
2025-12-04T09:45:15.8385422Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8385478Z Autotune Choices Stats:
2025-12-04T09:45:15.8386228Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.8386358Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8386474Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8386649Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8387264Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8387874Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8388496Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8389098Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8389700Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8390337Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8391011Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8391617Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8392241Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8392862Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8392992Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.8393035Z Autotune Choices Stats:
2025-12-04T09:45:15.8393799Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.8394034Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8394204Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8394495Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8395147Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8395795Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8396427Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8397049Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8397692Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8398325Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8398964Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8399613Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8400250Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8400910Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8401041Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.8401115Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8401161Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8401201Z unimplemented []
2025-12-04T09:45:15.8401264Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8401378Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8401963Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8402000Z graph_break []
2025-12-04T09:45:15.8402077Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8402118Z Autotune Choices Stats:
2025-12-04T09:45:15.8402862Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.8403021Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8403137Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8403302Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8403929Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8404538Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8405150Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8405771Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8406381Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8406985Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8407626Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8408250Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8408857Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8409468Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8409603Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.8409644Z Autotune Choices Stats:
2025-12-04T09:45:15.8410460Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.8410684Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8410849Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8411145Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8411781Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8412428Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8413058Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8413690Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8414334Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8414965Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8415594Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8416247Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8416894Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8417518Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8417653Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.8417731Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8417773Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8417815Z unimplemented []
2025-12-04T09:45:15.8417877Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8417980Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8418579Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8418622Z graph_break []
2025-12-04T09:45:15.8418695Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8418739Z Autotune Choices Stats:
2025-12-04T09:45:15.8419492Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.8419632Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8419749Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8419923Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8420564Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8421195Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8421801Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8422410Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8423039Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8423645Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8424260Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8424889Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8425508Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8426117Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8426248Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.8426293Z Autotune Choices Stats:
2025-12-04T09:45:15.8427055Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.8427290Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8427462Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8427740Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8428374Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8429025Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8429667Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8430295Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8430967Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8431612Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8432242Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8432875Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8433528Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8434174Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8434306Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.8434380Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8434424Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8434462Z unimplemented []
2025-12-04T09:45:15.8434526Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8434627Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8435206Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8435244Z graph_break []
2025-12-04T09:45:15.8435320Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8435361Z Autotune Choices Stats:
2025-12-04T09:45:15.8436130Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.8436262Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8436377Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8436542Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8437167Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8437790Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8438415Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8439023Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8439629Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8440255Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8440904Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8441536Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8442158Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8442777Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8442909Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.8442950Z Autotune Choices Stats:
2025-12-04T09:45:15.8443720Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.8443943Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8444112Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8444406Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8445043Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8445671Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8446320Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8446964Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8447599Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8448233Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8448871Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8449504Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8450139Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8450822Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8450955Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.8451032Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8451075Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8451144Z unimplemented []
2025-12-04T09:45:15.8451206Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8451309Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8451882Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8451922Z graph_break []
2025-12-04T09:45:15.8451996Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8452041Z Autotune Choices Stats:
2025-12-04T09:45:15.8452780Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.8452925Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8453043Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8453207Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8453825Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8454453Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8455079Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8455698Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8456308Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8456916Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8457534Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8458137Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8458762Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8459381Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8459513Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:15.8459558Z Autotune Choices Stats:
2025-12-04T09:45:15.8460335Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.8460594Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8460765Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8461045Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8461700Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8462330Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8462957Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8463607Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8464252Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8464883Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8465510Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8466153Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8466784Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8467425Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8467565Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:15.8467641Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8467685Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8467725Z unimplemented []
2025-12-04T09:45:15.8467785Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8467890Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8468489Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8468527Z graph_break []
2025-12-04T09:45:15.8468605Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8468645Z Autotune Choices Stats:
2025-12-04T09:45:15.8469400Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.8469531Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8469646Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8469809Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8470449Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8471058Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8471699Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8472321Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8472940Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8473546Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8474162Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8474782Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8475387Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8476006Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8476147Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:15.8476188Z Autotune Choices Stats:
2025-12-04T09:45:15.8476961Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:15.8477184Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8477349Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8477634Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8478289Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8478931Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8479558Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8480200Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8480871Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8481517Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8482148Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8482780Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8483422Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8484045Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8484190Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:15.8484265Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8484326Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8484363Z unimplemented []
2025-12-04T09:45:15.8484427Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8484528Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8485110Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8485149Z graph_break []
2025-12-04T09:45:15.8485222Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8485264Z Autotune Choices Stats:
2025-12-04T09:45:15.8486022Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:15.8486154Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8486272Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8486434Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8487048Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8487670Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8488283Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8488896Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8489514Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8490135Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8490771Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8491376Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8491998Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8492606Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8492752Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:15.8492793Z Autotune Choices Stats:
2025-12-04T09:45:15.8493568Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:15.8493808Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8493992Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8494273Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8494922Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8495551Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8496190Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8496821Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8497464Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8498110Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8498749Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8499380Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8500012Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8500689Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8500818Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:15.8500893Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8500938Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8500979Z unimplemented []
2025-12-04T09:45:15.8501039Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8501157Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8501737Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8501787Z graph_break []
2025-12-04T09:45:15.8501861Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8501902Z Autotune Choices Stats:
2025-12-04T09:45:15.8502673Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:15.8502802Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8502918Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8503081Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8503696Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8504327Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8504949Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8505553Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8506170Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8506790Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8507414Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8508020Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8508632Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8509247Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8509379Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:15.8509420Z Autotune Choices Stats:
2025-12-04T09:45:15.8510190Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.8510455Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8510627Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8510908Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8511558Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8512186Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8512813Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8513453Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8514082Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8514732Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8515373Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8516016Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8516648Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8517279Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8517425Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:15.8517500Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8517546Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8517583Z unimplemented []
2025-12-04T09:45:15.8517648Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8517749Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8518330Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8518381Z graph_break []
2025-12-04T09:45:15.8518455Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8518498Z Autotune Choices Stats:
2025-12-04T09:45:15.8519259Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:15.8519394Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8519510Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8519685Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8520302Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8520936Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8521561Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8522167Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8522778Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8523398Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8524020Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8524643Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8525257Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8525864Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8526007Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:15.8526051Z Autotune Choices Stats:
2025-12-04T09:45:15.8526816Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:15.8527049Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8527219Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8527505Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8528148Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8528790Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8529414Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8530044Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8530728Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8531366Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8532007Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8532650Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8533303Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8533934Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8534064Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:15.8534141Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8534183Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8534224Z unimplemented []
2025-12-04T09:45:15.8534284Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8534387Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8534986Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8535026Z graph_break []
2025-12-04T09:45:15.8535102Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8535142Z Autotune Choices Stats:
2025-12-04T09:45:15.8535886Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:15.8536041Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8536159Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8536323Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8536951Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8537559Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8538169Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8538784Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8539396Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8540005Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8540664Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8541286Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8541898Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8542512Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8542644Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:15.8542684Z Autotune Choices Stats:
2025-12-04T09:45:15.8543472Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:15.8543693Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8543862Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8544157Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8544789Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8545435Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8546074Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8546703Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8547354Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8547988Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8548620Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8549275Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8549915Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8550579Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8550713Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:15.8550787Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8550834Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8550874Z unimplemented []
2025-12-04T09:45:15.8550939Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8551043Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8551625Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8551686Z graph_break []
2025-12-04T09:45:15.8551761Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8551804Z Autotune Choices Stats:
2025-12-04T09:45:15.8552546Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:15.8552692Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8552808Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8552989Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8553608Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8554231Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8554845Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8555450Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8556070Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8556677Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8557293Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8557924Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8558548Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8559151Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8559286Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:15.8559329Z Autotune Choices Stats:
2025-12-04T09:45:15.8560104Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:15.8560340Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8560547Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8560829Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8561467Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8562125Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8562764Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8563391Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8564023Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8564662Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8565280Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8565909Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8566557Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8567206Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8567336Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:15.8567415Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8567457Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8567497Z unimplemented []
2025-12-04T09:45:15.8567558Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8567663Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8568248Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8568289Z graph_break []
2025-12-04T09:45:15.8568363Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8568406Z Autotune Choices Stats:
2025-12-04T09:45:15.8569165Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:15.8569296Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8569417Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8569577Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8570208Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8570860Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8571487Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8572091Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8572692Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8573307Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8573924Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8574534Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8575167Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8575789Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8575920Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:15.8575963Z Autotune Choices Stats:
2025-12-04T09:45:15.8576726Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.8576948Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8577118Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8577415Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8578046Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8578674Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8579324Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8579963Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8580630Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8581258Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8581904Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8582531Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8583159Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8583818Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8583952Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:15.8584027Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8584073Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8584115Z unimplemented []
2025-12-04T09:45:15.8584191Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8584295Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8584877Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8584916Z graph_break []
2025-12-04T09:45:15.8584993Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8585033Z Autotune Choices Stats:
2025-12-04T09:45:15.8585782Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:15.8585923Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8586038Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8586207Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8586826Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8587438Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8588064Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8588687Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8589299Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8589906Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8590566Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8591184Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8591814Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8592430Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8592563Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:15.8592604Z Autotune Choices Stats:
2025-12-04T09:45:15.8593375Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.8593598Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8593770Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8594051Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8594709Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8595339Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8595969Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8596615Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8597265Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8597899Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8598528Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8599169Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8599804Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8600492Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8600639Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:15.8600717Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8600759Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8600800Z unimplemented []
2025-12-04T09:45:15.8600862Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8600969Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8601565Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8601606Z graph_break []
2025-12-04T09:45:15.8601680Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8601725Z Autotune Choices Stats:
2025-12-04T09:45:15.8602478Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:15.8602608Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8602725Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8602885Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8603526Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8604134Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8604756Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8605378Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8605994Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8606602Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8607215Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8607842Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8608452Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8609071Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8609212Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:15.8609254Z Autotune Choices Stats:
2025-12-04T09:45:15.8610039Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:15.8610262Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8610464Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8610746Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8611388Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8612031Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8612656Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8613303Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8613954Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8614602Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8615233Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8615886Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8616534Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8617156Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8617299Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:15.8617394Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.8617444Z Traceback (most recent call last):
2025-12-04T09:45:15.8617614Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.8617658Z     self.assertTrue(
2025-12-04T09:45:15.8617765Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.8617818Z     raise self.failureException(msg)
2025-12-04T09:45:15.8617946Z AssertionError: False is not true : Log file /tmp/tmpukvz7181/flex_attention_configs.json was not created
2025-12-04T09:45:15.8617948Z 
2025-12-04T09:45:15.8618027Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.8618196Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.8618199Z 
2025-12-04T09:45:15.8618290Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.8618365Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8618411Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8618459Z unimplemented []
2025-12-04T09:45:15.8618524Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8619106Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.8619208Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8619248Z graph_break []
2025-12-04T09:45:15.8619322Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8619822Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.8619870Z   current_size = base.storage().size()
2025-12-04T09:45:15.8619913Z Autotune Choices Stats:
2025-12-04T09:45:15.8620696Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.8620830Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8620949Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8621109Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8621743Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8622362Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8622980Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8623589Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8624194Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8624810Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8625413Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8626024Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8626652Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8627271Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8627402Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.8627445Z Autotune Choices Stats:
2025-12-04T09:45:15.8628208Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.8628432Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8628601Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8628895Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8629531Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8630182Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8630854Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8631498Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8632136Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8632774Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8633425Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8634059Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8634687Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8635338Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8635470Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.8635547Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8635591Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8635630Z unimplemented []
2025-12-04T09:45:15.8635703Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8635805Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8636384Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8636422Z graph_break []
2025-12-04T09:45:15.8636498Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8636539Z Autotune Choices Stats:
2025-12-04T09:45:15.8637289Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.8637431Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8637546Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8637714Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8638331Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8638952Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8642656Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8643291Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8643889Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8644490Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8645114Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8645717Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8646341Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8646953Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8647086Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.8647126Z Autotune Choices Stats:
2025-12-04T09:45:15.8647910Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.8648134Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8648303Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8648586Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8649232Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8649858Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8650520Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8651176Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8651818Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8652447Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8653088Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8653732Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8654360Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8654996Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8655138Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.8655213Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8655256Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8655294Z unimplemented []
2025-12-04T09:45:15.8655354Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8655454Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8656040Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8656079Z graph_break []
2025-12-04T09:45:15.8656151Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8656191Z Autotune Choices Stats:
2025-12-04T09:45:15.8656939Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.8657070Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8657188Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8657351Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8657975Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8658583Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8659201Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8659813Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8660468Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8661080Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8661686Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8662299Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8662903Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8663528Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8663670Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.8663711Z Autotune Choices Stats:
2025-12-04T09:45:15.8664470Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.8664703Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8664871Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8665153Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8665800Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8666449Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8667076Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8667703Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8668357Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8669018Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8669647Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8670288Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8670991Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8671619Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8671761Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.8671835Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8671876Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8671928Z unimplemented []
2025-12-04T09:45:15.8671989Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8672091Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8672675Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8672711Z graph_break []
2025-12-04T09:45:15.8672785Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8672824Z Autotune Choices Stats:
2025-12-04T09:45:15.8673580Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.8673709Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8673826Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8673989Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8674605Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8675232Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8675839Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8676453Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8677073Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8677690Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8678298Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8678921Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8679540Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8680141Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8680282Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.8680321Z Autotune Choices Stats:
2025-12-04T09:45:15.8681112Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.8681348Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8681516Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8681812Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8682445Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8683074Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8683733Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8684360Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8685004Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8685642Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8686294Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8686919Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8687549Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8688183Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8688313Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.8688387Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8688429Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8688465Z unimplemented []
2025-12-04T09:45:15.8688526Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8688624Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8689217Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8689264Z graph_break []
2025-12-04T09:45:15.8689337Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8689376Z Autotune Choices Stats:
2025-12-04T09:45:15.8690123Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.8690253Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8690367Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8690569Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8691185Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8691790Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8692415Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8693021Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8693641Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8694262Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8694888Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8695497Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8696106Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8696738Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8696868Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.8696909Z Autotune Choices Stats:
2025-12-04T09:45:15.8697672Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.8697912Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8698078Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8698358Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8699010Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8699639Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8700264Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8700938Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8701572Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8702212Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8702849Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8703506Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8704135Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8704760Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8704898Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.8704971Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8705013Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8705050Z unimplemented []
2025-12-04T09:45:15.8705111Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8705212Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8705793Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8705841Z graph_break []
2025-12-04T09:45:15.8705915Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8705954Z Autotune Choices Stats:
2025-12-04T09:45:15.8706705Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.8706844Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8706959Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8707119Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8707747Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8708355Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8708957Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8709578Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8710183Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8710831Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8711458Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8712078Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8712685Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8713288Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8713435Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.8713475Z Autotune Choices Stats:
2025-12-04T09:45:15.8714240Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.8714461Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8714641Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8714937Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8715571Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8716206Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8716831Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8717461Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8718105Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8718732Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8719365Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8720014Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8720693Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8721325Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8721455Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.8721527Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8721568Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8721605Z unimplemented []
2025-12-04T09:45:15.8721666Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8721765Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8722359Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8722397Z graph_break []
2025-12-04T09:45:15.8722469Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8722509Z Autotune Choices Stats:
2025-12-04T09:45:15.8723252Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.8723406Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8723520Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8723681Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8724307Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8724913Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8725516Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8726135Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8726741Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8727346Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8727968Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8728586Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8729213Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8729819Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8729948Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.8729988Z Autotune Choices Stats:
2025-12-04T09:45:15.8730810Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.8731031Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8731196Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8731476Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8732134Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8732771Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8733412Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8734037Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8734669Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8735307Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8735934Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8736576Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8737214Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8737851Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8737979Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.8738054Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8738096Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8738133Z unimplemented []
2025-12-04T09:45:15.8738193Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8738296Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8738875Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8738912Z graph_break []
2025-12-04T09:45:15.8738995Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8739036Z Autotune Choices Stats:
2025-12-04T09:45:15.8739783Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.8739912Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8740039Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8740199Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8740842Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8741459Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8742069Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8742676Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8743298Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8743907Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8744542Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8745175Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8745796Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8746406Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8746536Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.8746577Z Autotune Choices Stats:
2025-12-04T09:45:15.8747338Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.8747557Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8747746Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8748024Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8748658Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8749296Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8749934Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8750617Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8751242Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8751890Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8752521Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8753148Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8753807Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8754457Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8754586Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.8754659Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8754701Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8754737Z unimplemented []
2025-12-04T09:45:15.8754797Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8754896Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8755469Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8755505Z graph_break []
2025-12-04T09:45:15.8755578Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8755616Z Autotune Choices Stats:
2025-12-04T09:45:15.8756368Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.8756498Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8756611Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8756774Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8757385Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8758013Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8758632Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8759238Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8759845Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8760502Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8761110Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8761725Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8762353Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8762981Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8763111Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.8763150Z Autotune Choices Stats:
2025-12-04T09:45:15.8763910Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.8764132Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8764298Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8764590Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8765225Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8765854Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8766505Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8767143Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8767778Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8768428Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8769070Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8769701Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8770332Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8771009Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8771138Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.8771213Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8771256Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8771295Z unimplemented []
2025-12-04T09:45:15.8771503Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8771628Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8772202Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8772241Z graph_break []
2025-12-04T09:45:15.8772314Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8772356Z Autotune Choices Stats:
2025-12-04T09:45:15.8773105Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.8773232Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8773365Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8773525Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8774143Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8774772Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8775398Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8776013Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8776619Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8777230Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8777865Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8778477Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8779086Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8779716Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8779846Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.8779887Z Autotune Choices Stats:
2025-12-04T09:45:15.8780705Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.8780925Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8781097Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8781378Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8782043Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8782671Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8783371Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8784027Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8784668Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8785302Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8785932Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8786567Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8787198Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8787829Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8787982Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.8788056Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8788101Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8788138Z unimplemented []
2025-12-04T09:45:15.8788199Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8788298Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8788887Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8788924Z graph_break []
2025-12-04T09:45:15.8788998Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8789038Z Autotune Choices Stats:
2025-12-04T09:45:15.8789778Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.8789909Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8790024Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8790187Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8790841Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8791449Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8792075Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8792693Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8793311Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8793918Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8794531Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8795163Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8795773Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8796385Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8796546Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.8796585Z Autotune Choices Stats:
2025-12-04T09:45:15.8797352Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.8797583Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8797749Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8798030Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8798659Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8799301Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8799925Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8800582Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8801235Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8801881Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8802507Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8803136Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8803776Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8804407Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8804561Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.8804637Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8804679Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8804719Z unimplemented []
2025-12-04T09:45:15.8804790Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8804896Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8805480Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8805520Z graph_break []
2025-12-04T09:45:15.8805594Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8805635Z Autotune Choices Stats:
2025-12-04T09:45:15.8806392Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.8806522Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8806639Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8806800Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8807419Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8808037Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8808645Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8809268Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8809886Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8810548Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8811156Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8811784Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8812412Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8813013Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8813157Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.8813199Z Autotune Choices Stats:
2025-12-04T09:45:15.8813960Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.8814193Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8814362Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8814650Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8815286Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8815919Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8816557Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8817182Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8817820Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8818461Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8819098Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8819732Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8820366Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8821038Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8821167Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.8821242Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8821285Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8821322Z unimplemented []
2025-12-04T09:45:15.8821383Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8821483Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8822079Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8822129Z graph_break []
2025-12-04T09:45:15.8822203Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8822243Z Autotune Choices Stats:
2025-12-04T09:45:15.8822987Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.8823131Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8823247Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8823409Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8824026Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8824627Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8825242Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8825849Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8826466Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8827087Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8827715Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8828327Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8828941Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8829557Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8829688Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:15.8829728Z Autotune Choices Stats:
2025-12-04T09:45:15.8830530Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.8830779Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8830945Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8831224Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8831875Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8832508Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8833154Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8833794Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8834430Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8835070Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8835708Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8836352Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8836981Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8837610Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8837741Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:15.8837824Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8837867Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8837904Z unimplemented []
2025-12-04T09:45:15.8837965Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8838069Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8838649Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8838698Z graph_break []
2025-12-04T09:45:15.8838772Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8838813Z Autotune Choices Stats:
2025-12-04T09:45:15.8839575Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.8839714Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8839832Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8839994Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8840654Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8841258Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8841864Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8842491Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8843096Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8843714Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8844347Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8844963Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8845568Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8846180Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8846309Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:15.8846360Z Autotune Choices Stats:
2025-12-04T09:45:15.8847119Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:15.8847339Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8847519Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8847807Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8848438Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8849079Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8849707Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8850338Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8851035Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8851665Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8852305Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8852949Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8853593Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8854219Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8854347Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:15.8854422Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8854463Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8854502Z unimplemented []
2025-12-04T09:45:15.8854562Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8854664Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8855251Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8855288Z graph_break []
2025-12-04T09:45:15.8855363Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8855401Z Autotune Choices Stats:
2025-12-04T09:45:15.8856147Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:15.8856296Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8856411Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8856572Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8857190Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8857809Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8858427Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8859040Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8859676Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8860281Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8860948Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8861569Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8862191Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8862800Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8862931Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:15.8862971Z Autotune Choices Stats:
2025-12-04T09:45:15.8863745Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:15.8863967Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8864137Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8864419Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8865066Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8865703Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8866360Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8866989Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8867618Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8868259Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8868884Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8869521Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8870153Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8870833Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8870962Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:15.8871037Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8871080Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8871117Z unimplemented []
2025-12-04T09:45:15.8871178Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8871279Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8871860Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8871899Z graph_break []
2025-12-04T09:45:15.8871987Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8872027Z Autotune Choices Stats:
2025-12-04T09:45:15.8872772Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:15.8872901Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8873030Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8873190Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8873823Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8874437Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8875044Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8875670Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8876297Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8876903Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8877517Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8878137Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8878760Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8879380Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8879510Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:15.8879552Z Autotune Choices Stats:
2025-12-04T09:45:15.8880306Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.8880557Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8880739Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8881018Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8881656Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8882298Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8882941Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8883581Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8884212Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8884844Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8885474Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8886105Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8886751Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8887386Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8887524Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:15.8887598Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8887639Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8887679Z unimplemented []
2025-12-04T09:45:15.8887738Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8887841Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8888419Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8888457Z graph_break []
2025-12-04T09:45:15.8888531Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8888570Z Autotune Choices Stats:
2025-12-04T09:45:15.8889326Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:15.8889455Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8889570Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8889731Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8890346Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8890991Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8891620Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8892225Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8892836Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8893456Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8894067Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8894694Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8895328Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8896020Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8896150Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:15.8896190Z Autotune Choices Stats:
2025-12-04T09:45:15.8896956Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:15.8897178Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8897347Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8897632Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8898280Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8898907Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8899544Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8900182Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8900848Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8901478Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8902119Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8902752Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8903400Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8904054Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8904183Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:15.8904257Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8904299Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8904339Z unimplemented []
2025-12-04T09:45:15.8904400Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8904511Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8905089Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8905127Z graph_break []
2025-12-04T09:45:15.8905201Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8905241Z Autotune Choices Stats:
2025-12-04T09:45:15.8905987Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:15.8906117Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8906243Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8906406Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8907026Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8907646Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8908278Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8908897Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8909506Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8910112Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8910767Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8911376Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8911982Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8912614Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8912744Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:15.8912786Z Autotune Choices Stats:
2025-12-04T09:45:15.8913564Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:15.8913784Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8913951Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8914235Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8914877Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8915575Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8916200Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8916849Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8917500Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8918137Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8918765Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8919417Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8920060Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8920721Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8920878Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:15.8920953Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8920994Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8921032Z unimplemented []
2025-12-04T09:45:15.8921092Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8921195Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8921789Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8921828Z graph_break []
2025-12-04T09:45:15.8921903Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8921942Z Autotune Choices Stats:
2025-12-04T09:45:15.8922691Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:15.8922820Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8922935Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8923096Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8923729Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8924336Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8924961Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8925576Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8926193Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8926804Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8927423Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8928050Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8928667Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8929295Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8929443Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:15.8929484Z Autotune Choices Stats:
2025-12-04T09:45:15.8930261Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:15.8930527Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8930695Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8930976Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8931616Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8932280Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8932912Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8933549Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8934210Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8934866Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8935494Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8936134Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8936782Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8937416Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8937546Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:15.8937632Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8937675Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8937713Z unimplemented []
2025-12-04T09:45:15.8937776Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8937888Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8938472Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8938510Z graph_break []
2025-12-04T09:45:15.8938587Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8938627Z Autotune Choices Stats:
2025-12-04T09:45:15.8939401Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:15.8939533Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8939650Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8939821Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8940471Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8941177Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8941801Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8942429Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8943053Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8943678Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8944290Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8944910Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8945539Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8946158Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8946299Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:15.8946339Z Autotune Choices Stats:
2025-12-04T09:45:15.8947106Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.8947339Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8947508Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8947805Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8948450Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8949084Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8949731Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8950367Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8951037Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8951700Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8952350Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8952987Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8953624Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8954280Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8954411Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:15.8954487Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8954528Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8954566Z unimplemented []
2025-12-04T09:45:15.8954626Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8954729Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8955325Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.8955377Z graph_break []
2025-12-04T09:45:15.8955450Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8955492Z Autotune Choices Stats:
2025-12-04T09:45:15.8956254Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:15.8956400Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8956517Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8956677Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8957304Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8957918Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8958544Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8959150Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8959773Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8960400Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8961060Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8961677Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8962294Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8962921Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8963054Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:15.8963095Z Autotune Choices Stats:
2025-12-04T09:45:15.8963870Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.8964119Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8964291Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8964575Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8965233Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8965870Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8966509Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8967154Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8967794Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8968444Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8969087Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8969740Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8970380Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8971035Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8971166Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:15.8971255Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8971298Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8971335Z unimplemented []
2025-12-04T09:45:15.8971396Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8971499Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8972078Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8972129Z graph_break []
2025-12-04T09:45:15.8972205Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8972245Z Autotune Choices Stats:
2025-12-04T09:45:15.8973006Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:15.8973156Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8973272Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8973437Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8974085Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8974705Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8975321Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8975952Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8976567Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8977186Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8977813Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8978436Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8979053Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8979684Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8979816Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:15.8979857Z Autotune Choices Stats:
2025-12-04T09:45:15.8980684Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:15.8980911Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8981091Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8981387Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8982038Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8982688Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8983318Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8983953Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8984612Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8985257Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8985904Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8986552Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8987205Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8987840Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8987974Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:15.8988049Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.8988090Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.8988130Z unimplemented []
2025-12-04T09:45:15.8988190Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.8988294Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.8988899Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.8988939Z graph_break []
2025-12-04T09:45:15.8989012Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.8989053Z Autotune Choices Stats:
2025-12-04T09:45:15.8989813Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:15.8989969Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8990086Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8990249Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8990903Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8991535Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8992151Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8992771Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8993420Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8994036Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.8994668Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8995303Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8995931Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8996547Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8996678Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:15.8996719Z Autotune Choices Stats:
2025-12-04T09:45:15.8997508Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.8997735Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.8997905Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.8998189Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.8998841Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.8999490Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9000139Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9000874Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9001517Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9002171Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9002805Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9003454Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9004104Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9004762Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9004894Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:15.9004989Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.9005038Z Traceback (most recent call last):
2025-12-04T09:45:15.9005195Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.9005239Z     self.assertTrue(
2025-12-04T09:45:15.9005347Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.9005398Z     raise self.failureException(msg)
2025-12-04T09:45:15.9005527Z AssertionError: False is not true : Log file /tmp/tmpu5r6r1gp/flex_attention_configs.json was not created
2025-12-04T09:45:15.9005530Z 
2025-12-04T09:45:15.9005607Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.9005777Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.9005780Z 
2025-12-04T09:45:15.9005886Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.9005963Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9006007Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9006044Z unimplemented []
2025-12-04T09:45:15.9006108Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9006692Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.9006793Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9006841Z graph_break []
2025-12-04T09:45:15.9006915Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9007420Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.9007479Z   current_size = base.storage().size()
2025-12-04T09:45:15.9007520Z Autotune Choices Stats:
2025-12-04T09:45:15.9008277Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.9008422Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9008541Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9008703Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9009329Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9009934Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9010594Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9011198Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9011815Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9012430Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9013055Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9013664Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9014274Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9014912Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9015042Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.9015083Z Autotune Choices Stats:
2025-12-04T09:45:15.9015848Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.9016089Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9016257Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9016537Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9017183Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9017815Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9018451Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9019084Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9019715Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9020353Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9021023Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9021677Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9022300Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9022929Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9023056Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.9023147Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9023189Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9023228Z unimplemented []
2025-12-04T09:45:15.9023288Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9023392Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9023969Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9024022Z graph_break []
2025-12-04T09:45:15.9024097Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9024138Z Autotune Choices Stats:
2025-12-04T09:45:15.9024888Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.9025030Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9025148Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9025311Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9025932Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9026534Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9027139Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9027756Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9028359Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9028975Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9029599Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9030213Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9030860Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9031461Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9031593Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.9031649Z Autotune Choices Stats:
2025-12-04T09:45:15.9032412Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.9032632Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9032812Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9033104Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9033741Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9034381Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9035005Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9035632Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9036274Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9036906Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9037544Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9038190Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9038843Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9039470Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9039601Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.9039676Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9039720Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9039758Z unimplemented []
2025-12-04T09:45:15.9039821Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9039920Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9040558Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9040597Z graph_break []
2025-12-04T09:45:15.9040669Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9040709Z Autotune Choices Stats:
2025-12-04T09:45:15.9041453Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.9041614Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9041730Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9041891Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9042505Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9043126Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9043735Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9044349Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9044965Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9045575Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9046195Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9046812Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9047429Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9048036Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9048167Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.9048208Z Autotune Choices Stats:
2025-12-04T09:45:15.9048978Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.9049201Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9049367Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9049644Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9050293Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9050953Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9051587Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9052214Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9052848Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9053492Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9054128Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9054783Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9055430Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9056078Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9056207Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.9056285Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9056326Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9056364Z unimplemented []
2025-12-04T09:45:15.9056425Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9056528Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9057104Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9057142Z graph_break []
2025-12-04T09:45:15.9057214Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9057265Z Autotune Choices Stats:
2025-12-04T09:45:15.9058010Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.9058137Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9058266Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9058428Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9059053Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9059662Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9060284Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9060927Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9061528Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9062162Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9062770Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9063397Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9064020Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9064654Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9064784Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.9064824Z Autotune Choices Stats:
2025-12-04T09:45:15.9065589Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.9065810Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9065990Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9066271Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9066904Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9067542Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9068179Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9068817Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9069445Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9070076Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9070761Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9071395Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9072045Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9072695Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9072839Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.9072914Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9072957Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9072993Z unimplemented []
2025-12-04T09:45:15.9073054Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9073153Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9073727Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9073767Z graph_break []
2025-12-04T09:45:15.9073840Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9073879Z Autotune Choices Stats:
2025-12-04T09:45:15.9074652Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.9074785Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9074900Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9075061Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9075683Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9076313Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9076941Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9077549Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9078159Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9078775Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9079388Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9079995Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9080636Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9081258Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9081404Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.9081445Z Autotune Choices Stats:
2025-12-04T09:45:15.9082219Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.9082440Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9082607Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9082883Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9083525Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9084152Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9084792Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9085432Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9086093Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9086732Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9087398Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9088030Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9088669Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9089310Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9089449Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.9089526Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9089568Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9089607Z unimplemented []
2025-12-04T09:45:15.9089666Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9089768Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9090362Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9090441Z graph_break []
2025-12-04T09:45:15.9090516Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9090556Z Autotune Choices Stats:
2025-12-04T09:45:15.9091313Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.9091441Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9091581Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9091741Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9092367Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9092976Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9093608Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9094228Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9094838Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9095452Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9096082Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9096691Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9097305Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9097939Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9098072Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.9098113Z Autotune Choices Stats:
2025-12-04T09:45:15.9098895Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.9099120Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9099288Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9099570Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9100210Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9100863Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9101489Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9102134Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9102794Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9103418Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9104054Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9104709Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9105341Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9105976Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9106131Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.9106204Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9106248Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9106285Z unimplemented []
2025-12-04T09:45:15.9106345Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9106445Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9107022Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9107071Z graph_break []
2025-12-04T09:45:15.9107144Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9107186Z Autotune Choices Stats:
2025-12-04T09:45:15.9107925Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.9108054Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9108168Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9108328Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9108953Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9109562Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9110168Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9113861Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9114480Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9115088Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9115699Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9116321Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9116934Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9117551Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9117708Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.9117748Z Autotune Choices Stats:
2025-12-04T09:45:15.9118519Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.9118747Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9118916Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9119201Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9119844Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9120510Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9121146Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9121779Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9122448Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9123099Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9123732Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9124379Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9125024Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9125659Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9125789Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.9125874Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9125917Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9125957Z unimplemented []
2025-12-04T09:45:15.9126016Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9126117Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9126709Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9126748Z graph_break []
2025-12-04T09:45:15.9126824Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9126865Z Autotune Choices Stats:
2025-12-04T09:45:15.9127631Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.9127761Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9127880Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9128042Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9128658Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9129285Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9129900Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9130556Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9131180Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9131808Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9132425Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9133039Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9133666Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9134281Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9134411Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.9134466Z Autotune Choices Stats:
2025-12-04T09:45:15.9135238Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.9135472Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9135642Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9135934Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9136577Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9137231Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9137901Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9138530Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9139186Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9139840Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9140524Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9141154Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9141794Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9142452Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9142585Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.9142661Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9142704Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9142741Z unimplemented []
2025-12-04T09:45:15.9142805Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9142904Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9143485Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9143551Z graph_break []
2025-12-04T09:45:15.9143625Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9143664Z Autotune Choices Stats:
2025-12-04T09:45:15.9144413Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.9144553Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9144669Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9144831Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9145462Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9146082Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9146719Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9147330Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9147951Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9148575Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9149210Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9149830Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9150474Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9151107Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9151240Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.9151279Z Autotune Choices Stats:
2025-12-04T09:45:15.9152053Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.9152292Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9152472Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9152757Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9153425Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9154060Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9156630Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9157281Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9157918Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9158564Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9159199Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9159846Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9160514Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9161140Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9161272Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.9161353Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9161414Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9161453Z unimplemented []
2025-12-04T09:45:15.9161515Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9161616Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9162191Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9162230Z graph_break []
2025-12-04T09:45:15.9162323Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9162364Z Autotune Choices Stats:
2025-12-04T09:45:15.9163107Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.9163247Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9163364Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9163528Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9164155Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9164762Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9165368Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9165984Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9166587Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9167204Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9167824Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9168446Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9169052Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9169675Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9169807Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.9169848Z Autotune Choices Stats:
2025-12-04T09:45:15.9170664Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.9170885Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9171067Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9171346Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9171997Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9172639Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9173259Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9173889Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9174529Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9175157Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9175792Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9176434Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9177074Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9177706Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9177836Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.9177910Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9177954Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9177991Z unimplemented []
2025-12-04T09:45:15.9178052Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9178151Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9178734Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9178773Z graph_break []
2025-12-04T09:45:15.9178846Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9178885Z Autotune Choices Stats:
2025-12-04T09:45:15.9179634Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.9179772Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9179897Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9180058Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9180703Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9181325Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9181935Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9182546Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9183166Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9183770Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9184388Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9185012Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9185645Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9186251Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9186382Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.9186422Z Autotune Choices Stats:
2025-12-04T09:45:15.9187197Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.9187416Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9187582Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9187859Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9188511Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9189147Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9189780Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9190437Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9191065Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9191711Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9192334Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9192986Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9193622Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9194264Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9194394Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.9194470Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9194512Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9194550Z unimplemented []
2025-12-04T09:45:15.9194610Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9194712Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9195288Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9195326Z graph_break []
2025-12-04T09:45:15.9195399Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9195438Z Autotune Choices Stats:
2025-12-04T09:45:15.9196188Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.9196321Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9196436Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9196608Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9197222Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9197832Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9198455Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9199063Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9199669Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9200303Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9200953Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9201573Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9202195Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9202822Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9202949Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.9202990Z Autotune Choices Stats:
2025-12-04T09:45:15.9203756Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.9203975Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9204162Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9204444Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9205083Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9205714Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9206349Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9207001Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9207632Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9208263Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9208904Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9209531Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9210164Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9210838Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9210981Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.9211056Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9211097Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9211135Z unimplemented []
2025-12-04T09:45:15.9211195Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9211295Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9211874Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9211912Z graph_break []
2025-12-04T09:45:15.9211985Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9212024Z Autotune Choices Stats:
2025-12-04T09:45:15.9212785Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.9212913Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9213028Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9213190Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9213812Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9214445Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9215060Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9215663Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9216272Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9216871Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9217489Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9218097Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9218712Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9219328Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9219475Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:15.9219514Z Autotune Choices Stats:
2025-12-04T09:45:15.9220286Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.9220545Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9220711Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9220990Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9221643Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9222273Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9222914Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9223548Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9224185Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9224815Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9225441Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9226082Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9226713Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9227349Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9227488Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:15.9227562Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9227605Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9227642Z unimplemented []
2025-12-04T09:45:15.9227703Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9227803Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9228398Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9228437Z graph_break []
2025-12-04T09:45:15.9228509Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9228550Z Autotune Choices Stats:
2025-12-04T09:45:15.9229299Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.9229427Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9229542Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9229716Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9230335Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9230985Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9231617Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9232241Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9232850Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9233454Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9234070Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9234686Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9235294Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9235909Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9236053Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:15.9236094Z Autotune Choices Stats:
2025-12-04T09:45:15.9236877Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:15.9237097Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9237265Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9237545Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9238192Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9238822Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9239446Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9240085Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9240745Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9241391Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9242015Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9242655Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9243279Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9243904Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9244057Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:15.9244131Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9244172Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9244209Z unimplemented []
2025-12-04T09:45:15.9244268Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9244368Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9244946Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9244983Z graph_break []
2025-12-04T09:45:15.9245066Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9245106Z Autotune Choices Stats:
2025-12-04T09:45:15.9245858Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:15.9245985Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9246102Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9246264Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9246897Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9247504Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9248115Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9248745Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9249363Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9249971Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9250639Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9251258Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9251865Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9252477Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9252638Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:15.9252677Z Autotune Choices Stats:
2025-12-04T09:45:15.9253441Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:15.9253677Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9253845Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9254123Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9254751Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9255388Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9256015Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9256639Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9257287Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9257944Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9258569Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9259200Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9259840Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9260502Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9260631Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:15.9260719Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9260761Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9260799Z unimplemented []
2025-12-04T09:45:15.9260860Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9260959Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9261550Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9261588Z graph_break []
2025-12-04T09:45:15.9261661Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9261702Z Autotune Choices Stats:
2025-12-04T09:45:15.9262452Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:15.9262580Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9262693Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9262858Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9263472Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9264090Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9264699Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9265325Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9265949Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9266564Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9267176Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9267803Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9268447Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9269054Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9269184Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:15.9269234Z Autotune Choices Stats:
2025-12-04T09:45:15.9269998Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.9270227Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9270395Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9270798Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9271432Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9272066Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9272721Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9273349Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9273986Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9274642Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9275280Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9275915Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9276543Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9277175Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9277306Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:15.9277379Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9277423Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9277461Z unimplemented []
2025-12-04T09:45:15.9277520Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9277621Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9278207Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9278267Z graph_break []
2025-12-04T09:45:15.9278342Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9278381Z Autotune Choices Stats:
2025-12-04T09:45:15.9279132Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:15.9279258Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9279385Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9279546Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9280168Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9280799Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9281423Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9282030Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9282647Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9283264Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9283906Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9284515Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9285120Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9285737Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9285870Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:15.9285910Z Autotune Choices Stats:
2025-12-04T09:45:15.9286673Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:15.9286903Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9287080Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9287360Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9288010Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9288642Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9289267Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9289924Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9290583Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9291209Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9291857Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9292499Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9293127Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9293754Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9293885Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:15.9293957Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9294001Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9294049Z unimplemented []
2025-12-04T09:45:15.9294110Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9294210Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9294798Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9294835Z graph_break []
2025-12-04T09:45:15.9294907Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9294958Z Autotune Choices Stats:
2025-12-04T09:45:15.9295703Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:15.9295840Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9295955Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9296122Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9296748Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9297349Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9297962Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9298574Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9299181Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9299798Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9300446Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9301065Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9301670Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9302276Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9302406Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:15.9302447Z Autotune Choices Stats:
2025-12-04T09:45:15.9303226Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:15.9303446Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9303614Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9303906Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9304550Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9305186Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9305815Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9306452Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9307109Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9307739Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9308385Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9309020Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9309669Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9310300Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9310450Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:15.9310527Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9310567Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9310606Z unimplemented []
2025-12-04T09:45:15.9310666Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9310766Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9311356Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9311396Z graph_break []
2025-12-04T09:45:15.9311470Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9311510Z Autotune Choices Stats:
2025-12-04T09:45:15.9312251Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:15.9312391Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9312521Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9312682Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9313298Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9313914Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9314523Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9315128Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9315742Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9316349Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9316975Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9317593Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9318210Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9318818Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9318950Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:15.9318992Z Autotune Choices Stats:
2025-12-04T09:45:15.9319772Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:15.9319991Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9320166Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9320476Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9321123Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9321760Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9322400Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9323025Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9323655Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9324300Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9324930Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9325569Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9326206Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9326847Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9326977Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:15.9327050Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9327094Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9327132Z unimplemented []
2025-12-04T09:45:15.9327191Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9327291Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9327868Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9327905Z graph_break []
2025-12-04T09:45:15.9327978Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9328019Z Autotune Choices Stats:
2025-12-04T09:45:15.9328773Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:15.9328901Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9329016Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9329188Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9329808Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9330454Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9331079Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9332956Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9333563Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9334211Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9334821Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9335448Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9336053Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9336663Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9336793Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:15.9336877Z Autotune Choices Stats:
2025-12-04T09:45:15.9337647Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.9337868Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9338046Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9338328Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9338963Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9339603Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9340231Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9340901Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9341536Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9342192Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9342832Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9343467Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9344110Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9344736Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9344869Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:15.9344956Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9345000Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9345038Z unimplemented []
2025-12-04T09:45:15.9345099Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9345199Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9345782Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9345832Z graph_break []
2025-12-04T09:45:15.9345906Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9345947Z Autotune Choices Stats:
2025-12-04T09:45:15.9346697Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:15.9346827Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9346945Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9347103Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9347711Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9348328Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9348937Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9349559Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9350174Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9350801Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9351424Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9352136Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9352758Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9353366Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9353507Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:15.9353549Z Autotune Choices Stats:
2025-12-04T09:45:15.9354305Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.9354539Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9354711Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9354991Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9355638Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9356268Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9356899Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9357524Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9358166Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9358797Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9359433Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9360071Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9360731Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9361379Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9361507Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:15.9361581Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9361624Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9361665Z unimplemented []
2025-12-04T09:45:15.9361730Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9361833Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9362422Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9362461Z graph_break []
2025-12-04T09:45:15.9362535Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9362588Z Autotune Choices Stats:
2025-12-04T09:45:15.9363341Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:15.9363468Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9363584Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9363759Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9364375Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9364987Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9365613Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9366230Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9366838Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9367455Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9368074Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9368681Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9369296Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9369910Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9370038Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:15.9370080Z Autotune Choices Stats:
2025-12-04T09:45:15.9370914Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:15.9371133Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9371319Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9371603Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9372238Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9372884Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9373514Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9374159Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9374791Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9375448Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9376087Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9376724Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9377377Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9378010Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9378149Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:15.9378224Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9378268Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9378305Z unimplemented []
2025-12-04T09:45:15.9378366Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9378465Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9379043Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9379084Z graph_break []
2025-12-04T09:45:15.9379167Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9379209Z Autotune Choices Stats:
2025-12-04T09:45:15.9379961Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:15.9380101Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9380219Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9380377Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9381024Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9381638Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9382249Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9382871Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9383494Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9384100Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9384730Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9385355Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9385972Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9386575Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9386718Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:15.9386759Z Autotune Choices Stats:
2025-12-04T09:45:15.9387517Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.9387739Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9387917Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9388194Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9388843Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9389486Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9390117Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9390767Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9391412Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9392060Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9392687Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9393334Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9393981Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9394613Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9394740Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:15.9394815Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9394868Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9394908Z unimplemented []
2025-12-04T09:45:15.9394968Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9395067Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9395637Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9395675Z graph_break []
2025-12-04T09:45:15.9395750Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9395791Z Autotune Choices Stats:
2025-12-04T09:45:15.9396560Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.9396688Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9396804Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9396986Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9397595Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9398216Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9398824Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9399445Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9400067Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9400728Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9401341Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9401963Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9402608Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9403211Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9403341Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:15.9403395Z Autotune Choices Stats:
2025-12-04T09:45:15.9404161Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.9404379Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9404548Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9404842Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9405476Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9406118Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9406751Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9407378Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9408022Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9408660Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9409293Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9409923Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9410600Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9411243Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9411374Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:15.9411466Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:15.9411513Z Traceback (most recent call last):
2025-12-04T09:45:15.9411668Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:15.9411713Z     self.assertTrue(
2025-12-04T09:45:15.9411820Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:15.9411872Z     raise self.failureException(msg)
2025-12-04T09:45:15.9412014Z AssertionError: False is not true : Log file /tmp/tmpzrs9u7ki/flex_attention_configs.json was not created
2025-12-04T09:45:15.9412018Z 
2025-12-04T09:45:15.9412097Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:15.9412264Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:15.9412268Z 
2025-12-04T09:45:15.9412357Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:15.9412432Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9412474Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9412511Z unimplemented []
2025-12-04T09:45:15.9412573Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9413154Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:15.9413273Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9413313Z graph_break []
2025-12-04T09:45:15.9413385Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9413887Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:15.9413950Z   current_size = base.storage().size()
2025-12-04T09:45:15.9413991Z Autotune Choices Stats:
2025-12-04T09:45:15.9414742Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.9414870Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9414989Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9415158Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9415774Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9416387Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9417002Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9417624Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9418234Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9418848Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9419467Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9420075Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9420719Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9421343Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9421472Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:15.9421515Z Autotune Choices Stats:
2025-12-04T09:45:15.9422286Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:15.9422505Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9422687Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9422970Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9423603Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9424246Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9424872Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9425512Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9426143Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9426785Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9427415Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9428051Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9428682Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9429306Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9429458Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:15.9429533Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9429577Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9429614Z unimplemented []
2025-12-04T09:45:15.9429675Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9429775Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9430362Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9430402Z graph_break []
2025-12-04T09:45:15.9430527Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9430569Z Autotune Choices Stats:
2025-12-04T09:45:15.9431323Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.9431465Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9431580Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9431742Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9432376Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9432980Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9433582Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9434205Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9434820Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9435421Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9436036Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9436654Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9437262Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9437865Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9438008Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:15.9438049Z Autotune Choices Stats:
2025-12-04T09:45:15.9438805Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.9439025Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9439200Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9439478Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9440121Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9440787Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9441488Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9442109Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9442755Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9443408Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9444033Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9444680Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9445328Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9445950Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9446080Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:15.9446156Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9446209Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9446249Z unimplemented []
2025-12-04T09:45:15.9446311Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9446412Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9446987Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9447025Z graph_break []
2025-12-04T09:45:15.9447099Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9447142Z Autotune Choices Stats:
2025-12-04T09:45:15.9447897Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:15.9448024Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9448140Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9448312Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9448929Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9449544Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9450153Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9450818Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9451436Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9452065Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9452667Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9453290Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9453914Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9454521Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9454650Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:15.9454703Z Autotune Choices Stats:
2025-12-04T09:45:15.9455460Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.9455677Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9455846Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9456135Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9456766Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9457403Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9458054Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9458681Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9459313Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9459949Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9460623Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9461247Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9461894Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9462561Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9462692Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:15.9462766Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9462810Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9462847Z unimplemented []
2025-12-04T09:45:15.9462909Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9463009Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9463583Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9463636Z graph_break []
2025-12-04T09:45:15.9463711Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9463751Z Autotune Choices Stats:
2025-12-04T09:45:15.9464495Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:15.9464624Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9464748Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9464911Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9465529Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9466147Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9466771Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9467376Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9467992Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9468599Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9469220Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9469830Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9470484Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9471105Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9471239Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:15.9471281Z Autotune Choices Stats:
2025-12-04T09:45:15.9472048Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.9472283Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9472452Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9472729Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9473382Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9474006Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9474652Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9475289Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9475921Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9476549Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9477190Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9477839Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9478467Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9479108Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9479238Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:15.9479314Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9479358Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9479397Z unimplemented []
2025-12-04T09:45:15.9479466Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9479567Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9480139Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9480178Z graph_break []
2025-12-04T09:45:15.9480250Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9480305Z Autotune Choices Stats:
2025-12-04T09:45:15.9481089Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:15.9481218Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9481333Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9481496Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9482134Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9482737Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9483357Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9483976Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9484584Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9485203Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9485812Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9486440Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9487046Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9487663Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9487792Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:15.9487836Z Autotune Choices Stats:
2025-12-04T09:45:15.9488610Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:15.9488830Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9488999Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9489299Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9489931Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9490616Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9491242Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9491879Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9492517Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9493149Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9493794Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9494429Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9495074Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9495698Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9495837Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:15.9495915Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9495959Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9495997Z unimplemented []
2025-12-04T09:45:15.9496059Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9496159Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9496751Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9496790Z graph_break []
2025-12-04T09:45:15.9496865Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9496905Z Autotune Choices Stats:
2025-12-04T09:45:15.9497654Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:15.9497793Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9497908Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9498072Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9498690Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9499305Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9499910Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9500562Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9501185Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9501785Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9502408Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9503031Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9503652Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9504258Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9504401Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:15.9504443Z Autotune Choices Stats:
2025-12-04T09:45:15.9505215Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.9505436Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9505602Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9505886Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9506540Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9507190Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9507845Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9508470Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9509110Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9509751Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9510371Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9511046Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9511674Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9512320Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9512450Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:15.9512522Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9512580Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9512617Z unimplemented []
2025-12-04T09:45:15.9512679Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9512780Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9513352Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9513392Z graph_break []
2025-12-04T09:45:15.9513464Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9513507Z Autotune Choices Stats:
2025-12-04T09:45:15.9514268Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:15.9514396Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9514512Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9514691Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9515315Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9515939Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9516570Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9517177Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9517791Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9518408Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9519022Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9519635Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9520237Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9520889Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9521018Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:15.9521072Z Autotune Choices Stats:
2025-12-04T09:45:15.9521833Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:15.9522053Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9522233Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9522512Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9523155Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9523797Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9524438Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9525081Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9525705Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9526343Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9526984Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9527615Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9528253Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9528877Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9529007Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:15.9529090Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9529132Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9529170Z unimplemented []
2025-12-04T09:45:15.9529229Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9529330Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9529906Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9529953Z graph_break []
2025-12-04T09:45:15.9530027Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9530068Z Autotune Choices Stats:
2025-12-04T09:45:15.9530854Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:15.9530986Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9531101Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9531261Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9531880Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9532497Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9533104Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9533722Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9534338Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9534941Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9535564Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9536170Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9536783Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9537392Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9537522Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:15.9537574Z Autotune Choices Stats:
2025-12-04T09:45:15.9538335Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:15.9538565Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9538732Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9539011Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9539648Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9540278Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9540949Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9541575Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9542224Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9542848Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9543494Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9544147Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9544772Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9545417Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9545547Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:15.9545620Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9545664Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9545701Z unimplemented []
2025-12-04T09:45:15.9545763Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9545861Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9546452Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9546490Z graph_break []
2025-12-04T09:45:15.9546563Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9546615Z Autotune Choices Stats:
2025-12-04T09:45:15.9547370Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:15.9547496Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9547612Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9547783Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9548400Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9549008Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9549626Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9550236Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9550892Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9551509Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9552117Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9552742Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9553349Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9553967Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9554099Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:15.9554141Z Autotune Choices Stats:
2025-12-04T09:45:15.9554916Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:15.9555135Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9555302Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9555591Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9556241Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9556883Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9557508Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9558155Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9558811Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9559470Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9560105Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9560782Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9561425Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9562054Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9562195Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:15.9562271Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9562313Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9562351Z unimplemented []
2025-12-04T09:45:15.9562416Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9562516Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9563090Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9563128Z graph_break []
2025-12-04T09:45:15.9563215Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9563257Z Autotune Choices Stats:
2025-12-04T09:45:15.9564007Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:15.9564148Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9564264Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9564422Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9565050Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9565657Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9566267Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9566876Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9567492Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9568099Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9568721Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9569338Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9569947Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9570578Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9570723Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:15.9570765Z Autotune Choices Stats:
2025-12-04T09:45:15.9571532Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.9571752Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9571931Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9572211Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9572858Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9573479Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9574124Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9574751Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9575389Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9576032Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9576690Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9577331Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9577969Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9578596Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9578725Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:15.9578799Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9578856Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9578894Z unimplemented []
2025-12-04T09:45:15.9578956Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9579055Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9579637Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9579674Z graph_break []
2025-12-04T09:45:15.9579746Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9579788Z Autotune Choices Stats:
2025-12-04T09:45:15.9580590Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:15.9580717Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9580831Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9581005Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9581626Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9582263Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9582875Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9583481Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9584101Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9584717Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9585324Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9585939Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9586558Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9587166Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9587294Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:15.9587334Z Autotune Choices Stats:
2025-12-04T09:45:15.9588105Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.9588324Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9588491Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9588770Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9589416Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9590057Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9590717Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9591343Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9591976Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9592615Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9593252Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9593881Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9594521Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9595162Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9595290Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:15.9595366Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9595408Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9595446Z unimplemented []
2025-12-04T09:45:15.9595505Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9595606Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9596177Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9596225Z graph_break []
2025-12-04T09:45:15.9596299Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9596340Z Autotune Choices Stats:
2025-12-04T09:45:15.9597088Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:15.9597216Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9597343Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9597502Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9598124Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9598746Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9599370Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9599976Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9600622Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9601249Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9601868Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9602480Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9603104Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9603720Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9603851Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:15.9603894Z Autotune Choices Stats:
2025-12-04T09:45:15.9604657Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.9604887Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9605058Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9605339Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9605987Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9606614Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9607256Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9607910Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9608541Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9609174Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9609809Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9610479Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9611111Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9611766Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9611894Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:15.9611967Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9612012Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9612049Z unimplemented []
2025-12-04T09:45:15.9612124Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9612225Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9612808Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9612847Z graph_break []
2025-12-04T09:45:15.9612922Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9612965Z Autotune Choices Stats:
2025-12-04T09:45:15.9613758Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.9613926Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9614041Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9614204Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9614848Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9615454Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9616071Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9616693Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9617296Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9617916Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9618529Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9619216Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9619835Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9620484Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9620614Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:15.9620655Z Autotune Choices Stats:
2025-12-04T09:45:15.9621467Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:15.9621687Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9621854Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9622157Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9622807Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9623452Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9624078Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9624717Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9625366Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9625992Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9626621Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9627264Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9627906Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9628531Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9628670Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:15.9628746Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9628787Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9628825Z unimplemented []
2025-12-04T09:45:15.9628885Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9628987Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9629583Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9629623Z graph_break []
2025-12-04T09:45:15.9629695Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9629738Z Autotune Choices Stats:
2025-12-04T09:45:15.9630528Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:15.9630675Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9630799Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9630959Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9631582Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9632300Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9632934Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9634246Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9635543Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9636824Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9638124Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9639390Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9640716Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9641966Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9642749Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:15.9642965Z Autotune Choices Stats:
2025-12-04T09:45:15.9651422Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:15.9652504Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9652955Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9653466Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9654451Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9655779Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9657212Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9658487Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9659787Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9670488Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9671763Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9673070Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9674349Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9675636Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9676420Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:15.9676667Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9676842Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9676954Z unimplemented []
2025-12-04T09:45:15.9677074Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9677278Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9677998Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9678634Z graph_break []
2025-12-04T09:45:15.9678764Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9678918Z Autotune Choices Stats:
2025-12-04T09:45:15.9679753Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:15.9680688Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9680969Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9681307Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9682119Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9683373Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9684657Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9685905Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9687163Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9688432Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9689682Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9690995Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9692247Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9693531Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9694299Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:15.9694505Z Autotune Choices Stats:
2025-12-04T09:45:15.9695352Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:15.9696364Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9696802Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9697288Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9698229Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9699529Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9700834Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9702123Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9703410Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9704715Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9706019Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9707306Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9708628Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9709926Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9710752Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:15.9711011Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9711166Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9711274Z unimplemented []
2025-12-04T09:45:15.9711391Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9711585Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9712301Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9712957Z graph_break []
2025-12-04T09:45:15.9713086Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9713238Z Autotune Choices Stats:
2025-12-04T09:45:15.9714083Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:15.9714985Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9715264Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9715574Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9716387Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9717651Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9718892Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9720170Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9721438Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9722720Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9723979Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9725221Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9726474Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9727719Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9728481Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:15.9728690Z Autotune Choices Stats:
2025-12-04T09:45:15.9729537Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.9730592Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9731012Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9731487Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9732449Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9733744Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9735035Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9736319Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9737631Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9738918Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9740211Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9741537Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9742837Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9744121Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9744905Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:15.9745142Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9745297Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9745404Z unimplemented []
2025-12-04T09:45:15.9745521Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9745714Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9746442Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9747083Z graph_break []
2025-12-04T09:45:15.9747210Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9747380Z Autotune Choices Stats:
2025-12-04T09:45:15.9748189Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:15.9749094Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9749375Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9749704Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9750551Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9751820Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9753102Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9754368Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9755656Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9756921Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9758189Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9759465Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9760757Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9762042Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9762816Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:15.9763020Z Autotune Choices Stats:
2025-12-04T09:45:15.9763869Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:15.9764886Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9765323Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9765835Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9766785Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9768108Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9769400Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9770751Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9772052Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9773401Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9774683Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9775979Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9777292Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9778581Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9779387Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:15.9779630Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9779786Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9779905Z unimplemented []
2025-12-04T09:45:15.9780022Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9780218Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9780959Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9781611Z graph_break []
2025-12-04T09:45:15.9781738Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9781904Z Autotune Choices Stats:
2025-12-04T09:45:15.9782716Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:15.9783629Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9783910Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9784222Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9785054Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9786302Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9787535Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9788798Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9790040Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9791339Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9792610Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9793870Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9795148Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9796389Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9797177Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:15.9797380Z Autotune Choices Stats:
2025-12-04T09:45:15.9798201Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:15.9799205Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9799636Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9800110Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9801099Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9802400Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9803698Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9804997Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9806300Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9807590Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9808890Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9810182Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9811505Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9812805Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9813591Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:15.9813828Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9813982Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9814107Z unimplemented []
2025-12-04T09:45:15.9814222Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9814417Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9815131Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9815769Z graph_break []
2025-12-04T09:45:15.9815898Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9816052Z Autotune Choices Stats:
2025-12-04T09:45:15.9816872Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:15.9817772Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9818052Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9818377Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9819198Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9820504Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9821752Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9823004Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9824267Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9825527Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9826773Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9828052Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9829312Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9830595Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9831360Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:15.9831566Z Autotune Choices Stats:
2025-12-04T09:45:15.9832400Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:15.9833426Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9833844Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9834320Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9835270Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9836574Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9837858Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9839158Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9840481Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9841784Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9843082Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9844371Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9845672Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9846964Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9847749Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:15.9847988Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9848144Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9848257Z unimplemented []
2025-12-04T09:45:15.9848374Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9848571Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9849294Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9849954Z graph_break []
2025-12-04T09:45:15.9850083Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9850235Z Autotune Choices Stats:
2025-12-04T09:45:15.9851094Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:15.9852014Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9852306Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9852620Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9853455Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9854745Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9856006Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9857249Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9858494Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9859751Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9861049Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9862303Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9863565Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9864843Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9865612Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:15.9865817Z Autotune Choices Stats:
2025-12-04T09:45:15.9866637Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:15.9867661Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9868079Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9868558Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9869512Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9870830Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9872130Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9873433Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9874718Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9876009Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9877317Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9878638Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9879931Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9881275Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9882066Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:15.9882303Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9882459Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9882569Z unimplemented []
2025-12-04T09:45:15.9882685Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9882897Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9883613Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:15.9884258Z graph_break []
2025-12-04T09:45:15.9884387Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9884541Z Autotune Choices Stats:
2025-12-04T09:45:15.9885375Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:15.9886275Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9886553Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9886867Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9887694Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9888934Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9890199Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9891501Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9892742Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9893990Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9895254Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9896517Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9897764Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9899013Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9899780Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:15.9899984Z Autotune Choices Stats:
2025-12-04T09:45:15.9900856Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:15.9901869Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9902287Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9902780Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9903721Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9905028Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9906319Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9907625Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9908926Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9910221Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9911539Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9912858Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9914166Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9915469Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9916265Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:15.9916503Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9916658Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9916767Z unimplemented []
2025-12-04T09:45:15.9916884Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9917079Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9917801Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9918447Z graph_break []
2025-12-04T09:45:15.9918576Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9918727Z Autotune Choices Stats:
2025-12-04T09:45:15.9919545Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:15.9920499Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9920778Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9921085Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9921903Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9923179Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9924423Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9925679Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9926944Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9928193Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9929451Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9930747Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9932026Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9933270Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9934060Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:15.9934264Z Autotune Choices Stats:
2025-12-04T09:45:15.9935085Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:15.9936129Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9936547Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9937028Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9937978Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9939300Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9940633Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9941907Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9943229Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9944544Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9945853Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9947151Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9948463Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9949769Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9950591Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:15.9950830Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9950997Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9951107Z unimplemented []
2025-12-04T09:45:15.9951224Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9951421Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9952135Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9952773Z graph_break []
2025-12-04T09:45:15.9952902Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9953055Z Autotune Choices Stats:
2025-12-04T09:45:15.9953869Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:15.9954776Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9955056Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9955383Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9956200Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9957447Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9958712Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9959960Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9961247Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9962521Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9963775Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9965044Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9966297Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9967564Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9968338Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:15.9968543Z Autotune Choices Stats:
2025-12-04T09:45:15.9969365Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:15.9970390Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9970849Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9971345Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9972299Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9973607Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9974912Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9976221Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9977532Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9978836Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9980141Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9981467Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9982792Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9984081Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9984876Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:15.9985120Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:15.9985285Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:15.9985395Z unimplemented []
2025-12-04T09:45:15.9985512Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:15.9985707Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:15.9986417Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:15.9987077Z graph_break []
2025-12-04T09:45:15.9987205Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:15.9987357Z Autotune Choices Stats:
2025-12-04T09:45:15.9988193Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:15.9989102Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:15.9989380Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:15.9989691Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:15.9990538Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9991815Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9993067Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9994337Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9995583Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9996841Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:15.9998124Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:15.9999383Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0000681Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0001940Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0002717Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.0002921Z Autotune Choices Stats:
2025-12-04T09:45:16.0003763Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.0004788Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0005208Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0005688Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0006654Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0007953Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0009261Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0010556Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0011871Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0013166Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0014467Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0015779Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0017072Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0018384Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0019174Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.0019411Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0019567Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0019676Z unimplemented []
2025-12-04T09:45:16.0019794Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0019988Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0020766Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0021406Z graph_break []
2025-12-04T09:45:16.0021534Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0021697Z Autotune Choices Stats:
2025-12-04T09:45:16.0022504Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.0023410Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0023686Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0024011Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0024822Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0026075Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0027349Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0028594Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0029849Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0031149Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0032407Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0033701Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0034948Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0036218Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0036987Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.0037191Z Autotune Choices Stats:
2025-12-04T09:45:16.0038023Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.0039035Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0039459Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0039957Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0040937Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0042266Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0043557Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0044862Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0046163Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0047479Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0048772Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0050079Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0051428Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0052726Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0053533Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.0053791Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.0053973Z Traceback (most recent call last):
2025-12-04T09:45:16.0054209Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.0054436Z     self.assertTrue(
2025-12-04T09:45:16.0054601Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.0054789Z     raise self.failureException(msg)
2025-12-04T09:45:16.0054997Z AssertionError: False is not true : Log file /tmp/tmpo_d0k7ct/flex_attention_configs.json was not created
2025-12-04T09:45:16.0055162Z 
2025-12-04T09:45:16.0055238Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.0055510Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.0055712Z 
2025-12-04T09:45:16.0055803Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.0056018Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0056172Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0056279Z unimplemented []
2025-12-04T09:45:16.0056397Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0057076Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.0057800Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0057972Z graph_break []
2025-12-04T09:45:16.0058101Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0058713Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.0059288Z   current_size = base.storage().size()
2025-12-04T09:45:16.0059410Z Autotune Choices Stats:
2025-12-04T09:45:16.0060252Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.0061191Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0061468Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0061798Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0062605Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0063853Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0065121Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0066361Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0067628Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0068891Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0070134Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0071431Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0072695Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0073959Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0074727Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.0074953Z Autotune Choices Stats:
2025-12-04T09:45:16.0075713Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.0075931Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0076115Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0076394Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0077029Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0077669Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0078301Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0078944Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0079566Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0080209Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0080884Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0081513Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0082155Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0082785Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0082917Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.0083004Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0083050Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0083088Z unimplemented []
2025-12-04T09:45:16.0083150Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0083250Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0083834Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0083891Z graph_break []
2025-12-04T09:45:16.0083966Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0084007Z Autotune Choices Stats:
2025-12-04T09:45:16.0084781Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.0084911Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0085027Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0085189Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0085805Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0086424Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0087030Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0087643Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0088248Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0088864Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0089486Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0090096Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0090766Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0091371Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0091505Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.0091545Z Autotune Choices Stats:
2025-12-04T09:45:16.0092331Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.0092564Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0092732Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0093009Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0093654Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0094280Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0094924Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0095546Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0096191Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0096820Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0097445Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0098086Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0098710Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0099341Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0099470Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.0099544Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0099589Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0099625Z unimplemented []
2025-12-04T09:45:16.0099688Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0099791Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0100381Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0100453Z graph_break []
2025-12-04T09:45:16.0100527Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0100589Z Autotune Choices Stats:
2025-12-04T09:45:16.0101335Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.0101465Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0101580Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0101754Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0102373Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0102991Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0103612Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0104218Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0104843Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0105466Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0106069Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0106685Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0107293Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0107909Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0108037Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.0108078Z Autotune Choices Stats:
2025-12-04T09:45:16.0108850Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.0109071Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0109237Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0109529Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0110167Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0110846Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0111469Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0112109Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0112757Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0113405Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0114038Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0114677Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0115322Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0115949Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0116093Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.0116169Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0116212Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0116251Z unimplemented []
2025-12-04T09:45:16.0116312Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0116414Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0117005Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0117044Z graph_break []
2025-12-04T09:45:16.0117117Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0117167Z Autotune Choices Stats:
2025-12-04T09:45:16.0117917Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.0118057Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0118172Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0118332Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0118968Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0119569Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0120178Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0120818Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0121424Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0122041Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0122673Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0123299Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0123937Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0124539Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0124688Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.0124727Z Autotune Choices Stats:
2025-12-04T09:45:16.0125493Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.0125714Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0125891Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0126171Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0126805Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0127447Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0128086Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0128713Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0129354Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0129984Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0130669Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0131317Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0131954Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0132610Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0132741Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.0132815Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0132860Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0132910Z unimplemented []
2025-12-04T09:45:16.0132972Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0133072Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0133653Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0133691Z graph_break []
2025-12-04T09:45:16.0133766Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0133808Z Autotune Choices Stats:
2025-12-04T09:45:16.0134568Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.0134696Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0134810Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0134980Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0135601Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0136231Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0136839Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0137441Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0138051Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0138665Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0139274Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0139900Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0140564Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0141169Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0141300Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.0141341Z Autotune Choices Stats:
2025-12-04T09:45:16.0142105Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.0142340Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0142506Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0142784Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0143434Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0144074Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0144701Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0145340Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0145973Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0146617Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0147245Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0147899Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0148545Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0149180Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0149309Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.0149383Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0149426Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0149465Z unimplemented []
2025-12-04T09:45:16.0149526Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0149627Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0150209Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0150258Z graph_break []
2025-12-04T09:45:16.0150333Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0150373Z Autotune Choices Stats:
2025-12-04T09:45:16.0151162Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.0151292Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0151432Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0151594Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0152210Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0152839Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0153474Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0154083Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0154688Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0155307Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0155936Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0156543Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0157164Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0157781Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0157912Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.0157952Z Autotune Choices Stats:
2025-12-04T09:45:16.0158718Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.0158948Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0159116Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0159397Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0160034Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0160702Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0161339Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0161980Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0162611Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0163241Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0163892Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0164541Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0165164Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0165805Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0165934Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.0166007Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0166051Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0166090Z unimplemented []
2025-12-04T09:45:16.0166152Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0166263Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0166842Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0166880Z graph_break []
2025-12-04T09:45:16.0166952Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0166993Z Autotune Choices Stats:
2025-12-04T09:45:16.0167745Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.0167882Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0167996Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0168158Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0168785Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0169393Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0170013Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0170669Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0171272Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0171882Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0172505Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0173126Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0173736Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0174359Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0174490Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.0174531Z Autotune Choices Stats:
2025-12-04T09:45:16.0175499Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.0178720Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0178900Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0179226Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0179873Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0180594Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0181221Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0181865Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0182512Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0183147Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0183775Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0184424Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0185068Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0185696Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0185842Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.0185922Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0185969Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0186009Z unimplemented []
2025-12-04T09:45:16.0186071Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0186175Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0186772Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0186813Z graph_break []
2025-12-04T09:45:16.0186890Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0186930Z Autotune Choices Stats:
2025-12-04T09:45:16.0187683Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.0187823Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0187942Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0188107Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0188724Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0189347Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0189957Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0190602Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0191228Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0191844Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0192467Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0193072Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0193696Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0194304Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0194448Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.0194488Z Autotune Choices Stats:
2025-12-04T09:45:16.0195245Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.0195482Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0195653Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0195935Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0196569Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0197205Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0197845Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0198463Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0199105Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0199749Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0200376Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0201061Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0201716Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0202357Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0202488Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.0202562Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0202621Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0202659Z unimplemented []
2025-12-04T09:45:16.0202721Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0202823Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0203400Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0203438Z graph_break []
2025-12-04T09:45:16.0203512Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0203554Z Autotune Choices Stats:
2025-12-04T09:45:16.0204313Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.0204444Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0204559Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0204722Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0205353Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0205976Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0206599Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0207204Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0207823Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0208438Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0209050Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0209672Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0210279Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0210929Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0211058Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.0211099Z Autotune Choices Stats:
2025-12-04T09:45:16.0211870Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.0212121Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0212287Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0212591Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0213219Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0213850Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0214491Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0215133Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0215765Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0216405Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0217044Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0217675Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0218322Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0218946Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0219077Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.0219151Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0219204Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0219243Z unimplemented []
2025-12-04T09:45:16.0219304Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0219403Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0219983Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0220033Z graph_break []
2025-12-04T09:45:16.0220106Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0220147Z Autotune Choices Stats:
2025-12-04T09:45:16.0220921Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.0221072Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0221189Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0221351Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0221969Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0222590Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0223197Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0223828Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0224437Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0225058Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0225690Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0226298Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0226910Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0227515Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0227646Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.0227686Z Autotune Choices Stats:
2025-12-04T09:45:16.0228469Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.0228699Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0228867Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0229147Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0229801Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0230472Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0231122Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0231750Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0232418Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0233048Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0233686Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0234336Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0234968Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0235606Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0235734Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.0235807Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0235850Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0235890Z unimplemented []
2025-12-04T09:45:16.0235952Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0236053Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0236655Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0236691Z graph_break []
2025-12-04T09:45:16.0236765Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0236806Z Autotune Choices Stats:
2025-12-04T09:45:16.0237559Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.0237688Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0237802Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0237965Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0238594Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0239196Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0239814Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0240461Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0241088Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0241693Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0242315Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0242938Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0243547Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0244172Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0244305Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.0244345Z Autotune Choices Stats:
2025-12-04T09:45:16.0245122Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.0245342Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0245507Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0245800Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0246435Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0247079Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0247703Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0248341Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0248978Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0249625Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0250252Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0250937Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0251585Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0252211Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0252358Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.0252432Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0252475Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0252514Z unimplemented []
2025-12-04T09:45:16.0252574Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0252674Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0253254Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0253293Z graph_break []
2025-12-04T09:45:16.0253366Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0253408Z Autotune Choices Stats:
2025-12-04T09:45:16.0254176Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.0254320Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0254435Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0254596Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0255226Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0255835Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0256448Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0257064Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0257670Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0258289Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0258907Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0259516Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0260143Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0260787Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0260932Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.0260973Z Autotune Choices Stats:
2025-12-04T09:45:16.0261741Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.0261964Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0262143Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0262424Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0263064Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0263703Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0264343Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0264973Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0265615Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0266261Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0266917Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0267550Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0268191Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0268833Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0268962Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.0269036Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0269078Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0269117Z unimplemented []
2025-12-04T09:45:16.0269188Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0269288Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0269860Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0269896Z graph_break []
2025-12-04T09:45:16.0269969Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0270009Z Autotune Choices Stats:
2025-12-04T09:45:16.0270909Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.0271039Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0271153Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0271325Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0271940Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0272567Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0273180Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0273789Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0274407Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0275016Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0275636Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0276262Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0276870Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0277507Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0277638Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.0277678Z Autotune Choices Stats:
2025-12-04T09:45:16.0278456Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.0278688Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0278855Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0279136Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0279777Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0280402Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0281089Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0281730Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0282359Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0283002Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0283627Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0284287Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0284920Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0285557Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0285689Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.0285763Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0285804Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0285844Z unimplemented []
2025-12-04T09:45:16.0285904Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0286004Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0286577Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0286626Z graph_break []
2025-12-04T09:45:16.0286699Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0286739Z Autotune Choices Stats:
2025-12-04T09:45:16.0287489Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.0287620Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0287736Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0287906Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0288521Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0289142Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0289761Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0290372Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0291007Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0291633Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0292266Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0292874Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0293497Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0294140Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0294271Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.0294312Z Autotune Choices Stats:
2025-12-04T09:45:16.0295068Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.0295302Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0295469Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0295745Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0296374Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0297016Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0297646Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0298289Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0298915Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0299550Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0300191Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0300909Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0301548Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0302187Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0302314Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.0302388Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0302431Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0302473Z unimplemented []
2025-12-04T09:45:16.0302534Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0302647Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0303227Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0303264Z graph_break []
2025-12-04T09:45:16.0303338Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0303377Z Autotune Choices Stats:
2025-12-04T09:45:16.0304123Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.0304266Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0304380Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0304541Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0305169Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0305778Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0306396Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0307016Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0307626Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0308232Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0308861Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0309483Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0310082Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0310722Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0310852Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.0310892Z Autotune Choices Stats:
2025-12-04T09:45:16.0311681Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.0311901Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0312067Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0312362Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0312995Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0313635Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0314261Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0314901Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0315563Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0316195Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0316822Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0317464Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0318123Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0318750Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0318891Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.0318966Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0319009Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0319048Z unimplemented []
2025-12-04T09:45:16.0319108Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0319208Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0319807Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0319848Z graph_break []
2025-12-04T09:45:16.0319920Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0319960Z Autotune Choices Stats:
2025-12-04T09:45:16.0320740Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.0320893Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0321008Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0321168Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0321791Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0322423Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0323032Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0323656Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0324282Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0324889Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0325501Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0326126Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0326749Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0327360Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0327499Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.0327541Z Autotune Choices Stats:
2025-12-04T09:45:16.0328296Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.0328528Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0328696Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0328976Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0329612Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0330253Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0330951Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0331575Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0332222Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0332874Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0333496Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0334127Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0334765Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0335408Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0335537Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.0335609Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0335652Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0335702Z unimplemented []
2025-12-04T09:45:16.0335763Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0335863Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0336446Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0336482Z graph_break []
2025-12-04T09:45:16.0336555Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0336595Z Autotune Choices Stats:
2025-12-04T09:45:16.0337353Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.0337482Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0337596Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0337756Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0338384Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0338991Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0339626Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0340234Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0340925Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0341548Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0342160Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0342792Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0343400Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0344027Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0344156Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.0344195Z Autotune Choices Stats:
2025-12-04T09:45:16.0344959Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.0345194Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0345363Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0345654Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0346291Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0346922Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0347557Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0348207Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0348834Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0349477Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0350110Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0350780Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0351438Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0352079Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0352209Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.0352283Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0352324Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0352377Z unimplemented []
2025-12-04T09:45:16.0352436Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0352538Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0353115Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0353168Z graph_break []
2025-12-04T09:45:16.0353242Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0353283Z Autotune Choices Stats:
2025-12-04T09:45:16.0354038Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.0354183Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0354299Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0354462Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0355078Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0355701Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0356307Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0356946Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0357554Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0358170Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0358790Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0359398Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0360014Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0360643Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0360774Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.0360814Z Autotune Choices Stats:
2025-12-04T09:45:16.0361593Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.0361826Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0361994Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0362273Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0362927Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0363561Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0364199Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0364823Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0365473Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0366106Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0366738Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0367385Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0368013Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0368652Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0368780Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.0368855Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0368897Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0368938Z unimplemented []
2025-12-04T09:45:16.0368999Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0369100Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0369684Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0369721Z graph_break []
2025-12-04T09:45:16.0369794Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0369834Z Autotune Choices Stats:
2025-12-04T09:45:16.0370622Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.0370772Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0370886Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0371049Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0371685Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0372294Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0372916Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0373521Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0374145Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0374745Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0375367Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0375994Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0376600Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0377213Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0377342Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.0377382Z Autotune Choices Stats:
2025-12-04T09:45:16.0378163Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.0378384Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0378551Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0378843Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0379489Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0380129Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0380791Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0381439Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0382071Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0382712Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0383332Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0383973Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0384629Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0385257Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0385400Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.0385474Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0385519Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0385557Z unimplemented []
2025-12-04T09:45:16.0385619Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0385718Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0386303Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0386343Z graph_break []
2025-12-04T09:45:16.0386415Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0386456Z Autotune Choices Stats:
2025-12-04T09:45:16.0387224Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.0387365Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0387483Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0387646Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0388276Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0388901Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0389506Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0390127Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0390776Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0391397Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0392009Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0392624Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0393253Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0393859Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0394011Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.0394053Z Autotune Choices Stats:
2025-12-04T09:45:16.0394822Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.0395045Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0395225Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0395502Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0396138Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0396777Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0397416Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0398050Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0398692Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0399325Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0399966Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0400638Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0401298Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0401954Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0402082Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.0402157Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0402198Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0402238Z unimplemented []
2025-12-04T09:45:16.0402300Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0402423Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0403002Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0403039Z graph_break []
2025-12-04T09:45:16.0403113Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0403153Z Autotune Choices Stats:
2025-12-04T09:45:16.0403919Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.0404051Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0404167Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0404330Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0404965Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0405579Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0406193Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0406795Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0407416Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0408025Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0408652Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0409268Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0409888Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0410552Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0410683Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.0410722Z Autotune Choices Stats:
2025-12-04T09:45:16.0411488Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.0411722Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0411888Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0412170Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0412809Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0413435Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0414073Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0414712Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0415340Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0415977Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0416602Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0417245Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0417883Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0418527Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0418667Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.0418741Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0418783Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0418821Z unimplemented []
2025-12-04T09:45:16.0418883Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0418984Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0419565Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0419614Z graph_break []
2025-12-04T09:45:16.0419687Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0419727Z Autotune Choices Stats:
2025-12-04T09:45:16.0420510Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.0420641Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0420757Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0420943Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0421557Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0422189Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0422816Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0423424Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0424024Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0424646Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0425258Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0425878Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0426501Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0427106Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0427248Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.0427289Z Autotune Choices Stats:
2025-12-04T09:45:16.0428057Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.0428289Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0428459Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0428741Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0429383Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0430045Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0430772Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0431403Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0432055Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0432689Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0433329Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0433963Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0434632Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0435272Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0435404Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.0435479Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0435521Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0435561Z unimplemented []
2025-12-04T09:45:16.0435624Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0435742Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0436319Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0436356Z graph_break []
2025-12-04T09:45:16.0436430Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0436470Z Autotune Choices Stats:
2025-12-04T09:45:16.0437221Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.0437370Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0437488Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0437649Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0438276Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0438893Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0439512Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0440127Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0440778Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0441386Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0442019Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0442657Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0443266Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0443888Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0444018Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.0444057Z Autotune Choices Stats:
2025-12-04T09:45:16.0444840Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.0445059Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0445226Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0445517Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0446151Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0446777Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0447416Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0448048Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0448691Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0449328Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0449952Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0450631Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0451280Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0451910Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0452054Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.0452131Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0452175Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0452214Z unimplemented []
2025-12-04T09:45:16.0452275Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0452374Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0452972Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0453012Z graph_break []
2025-12-04T09:45:16.0453085Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0453126Z Autotune Choices Stats:
2025-12-04T09:45:16.0453876Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.0454033Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0454146Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0454308Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0454935Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0455559Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0456172Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0456792Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0457427Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0458034Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0458658Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0459279Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0459914Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0460550Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0460702Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.0460742Z Autotune Choices Stats:
2025-12-04T09:45:16.0461514Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.0461749Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0461917Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0462197Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0462844Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0463488Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0464129Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0464749Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0465397Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0466037Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0466664Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0467300Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0467941Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0468578Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0468708Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.0468783Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0468825Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0468884Z unimplemented []
2025-12-04T09:45:16.0468945Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0469048Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0469637Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0469674Z graph_break []
2025-12-04T09:45:16.0469749Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0469788Z Autotune Choices Stats:
2025-12-04T09:45:16.0470579Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.0470708Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0470824Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0470986Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0471618Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0472227Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0472871Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0473479Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0474097Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0474714Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0475323Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0475943Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0476553Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0477173Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0477303Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.0477344Z Autotune Choices Stats:
2025-12-04T09:45:16.0478114Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.0478342Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0478509Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0478804Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0479448Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0480095Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0480781Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0481437Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0482066Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0482708Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0483351Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0483991Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0484638Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0485257Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0485389Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.0485463Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0485507Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0485561Z unimplemented []
2025-12-04T09:45:16.0485622Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0485722Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0486292Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0486342Z graph_break []
2025-12-04T09:45:16.0486417Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0486458Z Autotune Choices Stats:
2025-12-04T09:45:16.0487208Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.0487348Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0487464Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0487625Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0488245Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0488869Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0489485Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0490103Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0490752Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0491377Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0491999Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0492607Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0493233Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0493857Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0493988Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.0494029Z Autotune Choices Stats:
2025-12-04T09:45:16.0494816Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.0495047Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0495217Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0495498Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0496149Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0496775Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0497419Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0498048Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0498698Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0499331Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0499970Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0500664Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0501299Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0501950Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0502077Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.0502170Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.0502217Z Traceback (most recent call last):
2025-12-04T09:45:16.0502378Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.0502419Z     self.assertTrue(
2025-12-04T09:45:16.0502530Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.0502579Z     raise self.failureException(msg)
2025-12-04T09:45:16.0502722Z AssertionError: False is not true : Log file /tmp/tmpfx5o9jqp/flex_attention_configs.json was not created
2025-12-04T09:45:16.0502726Z 
2025-12-04T09:45:16.0502802Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.0502971Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.0502973Z 
2025-12-04T09:45:16.0503063Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.0503141Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0503199Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0503237Z unimplemented []
2025-12-04T09:45:16.0503299Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0503880Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.0503979Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0504016Z graph_break []
2025-12-04T09:45:16.0504090Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0504596Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.0504647Z   current_size = base.storage().size()
2025-12-04T09:45:16.0504688Z Autotune Choices Stats:
2025-12-04T09:45:16.0505441Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.0505583Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0505700Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0505863Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0506482Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0507096Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0507701Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0508319Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0508928Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0509536Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0510156Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0510800Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0511424Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0512028Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0512173Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.0512214Z Autotune Choices Stats:
2025-12-04T09:45:16.0512984Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.0513219Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0513388Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0513668Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0514299Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0514937Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0515580Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0516205Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0516844Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0517489Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0518118Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0518764Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0519406Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0520052Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0520182Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.0520257Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0520299Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0520349Z unimplemented []
2025-12-04T09:45:16.0520414Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0520535Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0521113Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0521151Z graph_break []
2025-12-04T09:45:16.0521224Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0521265Z Autotune Choices Stats:
2025-12-04T09:45:16.0522039Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.0522168Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0522285Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0522447Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0523087Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0523695Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0524330Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0524934Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0525565Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0526181Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0526788Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0527406Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0528023Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0528638Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0528767Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.0528808Z Autotune Choices Stats:
2025-12-04T09:45:16.0529592Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.0529826Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0529994Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0530288Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0530959Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0531594Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0532240Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0532882Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0533514Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0534160Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0534798Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0535430Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0536073Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0536700Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0536830Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.0536903Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0536945Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0536995Z unimplemented []
2025-12-04T09:45:16.0537057Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0537158Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0537742Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0537792Z graph_break []
2025-12-04T09:45:16.0537868Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0537909Z Autotune Choices Stats:
2025-12-04T09:45:16.0538654Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.0538795Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0538909Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0539072Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0539686Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0540303Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0540955Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0541581Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0542177Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0542798Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0543432Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0544043Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0544666Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0545271Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0545404Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.0545445Z Autotune Choices Stats:
2025-12-04T09:45:16.0546218Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.0546451Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0546618Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0546896Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0547550Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0548177Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0548816Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0549442Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0550093Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0550755Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0551400Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0552056Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0552689Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0553329Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0553458Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.0553532Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0553574Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0553613Z unimplemented []
2025-12-04T09:45:16.0553674Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0553775Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0554367Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0554406Z graph_break []
2025-12-04T09:45:16.0554479Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0554519Z Autotune Choices Stats:
2025-12-04T09:45:16.0555264Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.0555403Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0555520Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0555681Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0556313Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0556919Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0557545Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0558168Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0558786Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0559399Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0560020Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0560693Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0561314Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0561943Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0562074Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.0562117Z Autotune Choices Stats:
2025-12-04T09:45:16.0562909Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.0563133Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0563303Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0563605Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0564242Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0564884Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0565513Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0566163Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0566803Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0567448Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0568076Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0568717Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0569368Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0569997Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0570139Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.0570217Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0570259Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0570300Z unimplemented []
2025-12-04T09:45:16.0570361Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0570495Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0571075Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0571114Z graph_break []
2025-12-04T09:45:16.0571191Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0571233Z Autotune Choices Stats:
2025-12-04T09:45:16.0572006Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.0572150Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0572266Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0572436Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0573058Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0573680Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0574291Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0574915Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0575525Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0576147Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0576759Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0577383Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0578002Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0578606Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0578746Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.0578786Z Autotune Choices Stats:
2025-12-04T09:45:16.0579548Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.0579769Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0579945Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0580222Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0580888Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0581538Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0582185Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0582808Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0583457Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0584085Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0584739Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0585371Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0586008Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0586647Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0586775Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.0586849Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0586892Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0586930Z unimplemented []
2025-12-04T09:45:16.0587004Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0587104Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0587687Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0587724Z graph_break []
2025-12-04T09:45:16.0587798Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0587838Z Autotune Choices Stats:
2025-12-04T09:45:16.0588599Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.0588730Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0588844Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0589011Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0589639Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0590254Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0590900Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0591509Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0592127Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0592731Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0593354Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0593975Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0594582Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0595214Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0595344Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.0595385Z Autotune Choices Stats:
2025-12-04T09:45:16.0596139Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.0596370Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0596537Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0596820Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0597473Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0598102Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0598738Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0599381Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0600005Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0600662Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0601288Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0601936Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0602576Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0603209Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0603355Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.0603431Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0603472Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0603512Z unimplemented []
2025-12-04T09:45:16.0603572Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0603672Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0604247Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0604301Z graph_break []
2025-12-04T09:45:16.0604376Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0604415Z Autotune Choices Stats:
2025-12-04T09:45:16.0605166Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.0605294Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0605409Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0605586Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0606193Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0606811Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0607431Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0608035Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0608635Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0609255Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0609875Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0610531Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0611163Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0611768Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0611916Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.0611956Z Autotune Choices Stats:
2025-12-04T09:45:16.0612726Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.0612968Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0613138Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0613418Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0614053Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0614706Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0615341Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0615968Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0616613Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0617247Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0617885Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0618521Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0619167Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0619804Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0619934Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.0620006Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0620048Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0620085Z unimplemented []
2025-12-04T09:45:16.0620147Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0620257Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0620863Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0620901Z graph_break []
2025-12-04T09:45:16.0620974Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0621015Z Autotune Choices Stats:
2025-12-04T09:45:16.0621763Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.0621909Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0622023Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0622187Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0622831Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0623430Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0625167Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0625801Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0626420Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0627025Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0627653Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0628269Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0628878Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0629494Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0629664Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.0629704Z Autotune Choices Stats:
2025-12-04T09:45:16.0630534Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.0630758Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0630923Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0631224Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0631862Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0632495Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0633126Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0633767Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0634434Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0635064Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0635692Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0636342Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0636971Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0637596Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0637740Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.0637816Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0637857Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0637896Z unimplemented []
2025-12-04T09:45:16.0637957Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0638058Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0638669Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0638710Z graph_break []
2025-12-04T09:45:16.0638783Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0638824Z Autotune Choices Stats:
2025-12-04T09:45:16.0639571Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.0639711Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0639825Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0639984Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0640647Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0641254Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0641875Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0642518Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0643137Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0643745Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0644355Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0644980Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0645588Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0646192Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0646333Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.0646374Z Autotune Choices Stats:
2025-12-04T09:45:16.0647164Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.0647395Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0647564Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0647842Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0648477Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0649121Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0649745Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0650393Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0651080Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0651762Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0652388Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0653023Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0653673Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0654298Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0654428Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.0654503Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0654546Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0654583Z unimplemented []
2025-12-04T09:45:16.0654665Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0654768Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0655349Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0655385Z graph_break []
2025-12-04T09:45:16.0655474Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0655514Z Autotune Choices Stats:
2025-12-04T09:45:16.0656275Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.0656405Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0656521Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0656686Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0657312Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0657923Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0658535Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0659147Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0659766Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0660381Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0661030Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0661635Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0662263Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0662869Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0663001Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.0663042Z Autotune Choices Stats:
2025-12-04T09:45:16.0663812Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.0664069Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0664237Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0664536Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0665177Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0665804Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0666443Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0667076Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0667717Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0668367Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0669009Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0669639Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0670267Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0670932Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0671063Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.0671138Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0671179Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0671218Z unimplemented []
2025-12-04T09:45:16.0671278Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0671379Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0671953Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0672008Z graph_break []
2025-12-04T09:45:16.0672083Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0672125Z Autotune Choices Stats:
2025-12-04T09:45:16.0672907Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.0673059Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0673176Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0673337Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0673954Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0674588Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0675195Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0675809Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0676415Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0677048Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0677662Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0678271Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0678900Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0679509Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0679640Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.0679680Z Autotune Choices Stats:
2025-12-04T09:45:16.0680462Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.0680707Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0680875Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0681177Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0681832Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0682465Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0683092Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0683752Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0684384Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0685011Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0685662Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0686305Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0686930Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0687563Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0687691Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.0690189Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0690239Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0690279Z unimplemented []
2025-12-04T09:45:16.0690344Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0690498Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0691077Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0691115Z graph_break []
2025-12-04T09:45:16.0691193Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0691235Z Autotune Choices Stats:
2025-12-04T09:45:16.0691987Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.0692155Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0692300Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0692464Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0693099Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0693704Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0694327Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0694949Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0695559Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0696158Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0696788Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0697420Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0698026Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0698638Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0698770Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.0698812Z Autotune Choices Stats:
2025-12-04T09:45:16.0699585Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.0699808Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0699978Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0700269Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0700966Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0701609Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0702228Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0702866Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0703495Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0704127Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0704751Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0705410Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0706056Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0706687Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0706833Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.0706908Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0706953Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0706991Z unimplemented []
2025-12-04T09:45:16.0707054Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0707155Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0707734Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0707774Z graph_break []
2025-12-04T09:45:16.0707847Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0707888Z Autotune Choices Stats:
2025-12-04T09:45:16.0708633Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.0708776Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0708893Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0709054Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0709688Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0710314Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0710962Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0711594Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0712198Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0712800Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0713411Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0714068Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0714689Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0715292Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0715434Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.0715475Z Autotune Choices Stats:
2025-12-04T09:45:16.0716245Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.0716467Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0716635Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0716913Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0717558Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0718200Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0718977Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0719604Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0720245Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0720926Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0721549Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0722182Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0722833Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0723469Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0723598Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.0723676Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0723717Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0723756Z unimplemented []
2025-12-04T09:45:16.0723816Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0723931Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0724504Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0724541Z graph_break []
2025-12-04T09:45:16.0724616Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0724656Z Autotune Choices Stats:
2025-12-04T09:45:16.0725411Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.0725543Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0725661Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0725825Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0726445Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0727058Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0727681Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0728286Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0728900Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0729507Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0730124Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0730785Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0731434Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0732053Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0732183Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.0732224Z Autotune Choices Stats:
2025-12-04T09:45:16.0732994Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.0733228Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0733396Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0733685Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0734327Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0734952Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0735597Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0736238Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0736865Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0737505Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0738149Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0738797Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0739423Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0740076Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0740220Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.0740293Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0740335Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0740372Z unimplemented []
2025-12-04T09:45:16.0740476Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0740576Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0741150Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0741205Z graph_break []
2025-12-04T09:45:16.0741278Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0741319Z Autotune Choices Stats:
2025-12-04T09:45:16.0742065Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.0742195Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0742309Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0742471Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0743088Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0743707Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0744340Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0744944Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0745567Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0746201Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0746811Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0747424Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0748045Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0748678Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0748819Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.0748860Z Autotune Choices Stats:
2025-12-04T09:45:16.0749632Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.0749865Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0750033Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0750316Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0750999Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0751629Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0752257Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0752912Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0753566Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0754191Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0754829Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0755459Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0756092Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0756737Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0756865Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.0756958Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0757001Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0757039Z unimplemented []
2025-12-04T09:45:16.0757098Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0757200Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0757783Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0757822Z graph_break []
2025-12-04T09:45:16.0757896Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0757937Z Autotune Choices Stats:
2025-12-04T09:45:16.0758685Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.0758824Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0758940Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0759103Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0759720Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0760330Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0760975Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0761621Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0762229Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0762848Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0763470Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0764075Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0764688Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0765310Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0765451Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.0765492Z Autotune Choices Stats:
2025-12-04T09:45:16.0766264Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.0766485Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0766652Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0766943Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0767580Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0768211Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0768831Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0769468Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0770121Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0770786Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0771415Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0772066Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0772699Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0773326Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0773470Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.0773543Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0773586Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0773623Z unimplemented []
2025-12-04T09:45:16.0773684Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0773786Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0774387Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0774439Z graph_break []
2025-12-04T09:45:16.0774513Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0774554Z Autotune Choices Stats:
2025-12-04T09:45:16.0775297Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.0775437Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0775552Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0775714Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0776335Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0776939Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0777547Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0778182Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0778801Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0779407Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0780024Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0780687Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0781296Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0781893Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0782044Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.0782084Z Autotune Choices Stats:
2025-12-04T09:45:16.0782866Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.0783099Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0783264Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0783545Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0784184Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0784829Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0785454Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0786081Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0786725Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0787378Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0788004Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0788636Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0789275Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0789902Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0790031Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.0790105Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0790146Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0790184Z unimplemented []
2025-12-04T09:45:16.0790244Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0790357Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0790970Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0791008Z graph_break []
2025-12-04T09:45:16.0791097Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0791138Z Autotune Choices Stats:
2025-12-04T09:45:16.0791900Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.0792028Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0792144Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0792303Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0792939Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0793548Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0794160Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0794772Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0795403Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0796022Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0796638Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0797246Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0797865Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0798473Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0798603Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.0798644Z Autotune Choices Stats:
2025-12-04T09:45:16.0799409Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.0799636Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0799813Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0800105Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0800781Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0801409Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0802051Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0802694Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0803343Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0803986Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0804640Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0805276Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0805905Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0806543Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0806674Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.0806747Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0806789Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0806827Z unimplemented []
2025-12-04T09:45:16.0806888Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0806989Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0807572Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0807621Z graph_break []
2025-12-04T09:45:16.0807694Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0807734Z Autotune Choices Stats:
2025-12-04T09:45:16.0808491Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.0808622Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0808748Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0808907Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0809522Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0810141Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0810778Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0811385Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0811994Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0812619Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0813242Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0813851Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0814462Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0815086Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0815218Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.0815258Z Autotune Choices Stats:
2025-12-04T09:45:16.0816022Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.0816252Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0816418Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0816703Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0817361Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0817990Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0818618Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0819256Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0819889Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0820560Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0821213Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0821859Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0822487Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0823120Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0823265Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.0823340Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0823382Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0823420Z unimplemented []
2025-12-04T09:45:16.0823480Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0823581Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0824164Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0824201Z graph_break []
2025-12-04T09:45:16.0824275Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0824316Z Autotune Choices Stats:
2025-12-04T09:45:16.0825071Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.0825209Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0825334Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0825495Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0826133Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0826737Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0827352Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0827973Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0828601Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0829204Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0829832Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0830501Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0831114Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0831727Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0831855Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.0831898Z Autotune Choices Stats:
2025-12-04T09:45:16.0832666Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.0832886Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0833055Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0833348Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0834000Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0834641Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0835273Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0835897Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0836537Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0837173Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0837799Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0838454Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0839091Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0839723Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0839864Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.0839938Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0839981Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0840018Z unimplemented []
2025-12-04T09:45:16.0840080Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0840182Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0840797Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0840836Z graph_break []
2025-12-04T09:45:16.0840910Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0840950Z Autotune Choices Stats:
2025-12-04T09:45:16.0841699Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.0841848Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0841964Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0842126Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0842763Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0843386Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0843998Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0844617Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0845238Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0845864Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0846491Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0847121Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0847742Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0848351Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0848493Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.0848533Z Autotune Choices Stats:
2025-12-04T09:45:16.0849296Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.0849518Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0849684Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0849962Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0850629Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0851293Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0851932Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0852555Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0853209Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0853843Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0854471Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0855105Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0855761Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0856394Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0856525Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.0856600Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0856642Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0856681Z unimplemented []
2025-12-04T09:45:16.0856741Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0856842Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0857433Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0857471Z graph_break []
2025-12-04T09:45:16.0857545Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0857586Z Autotune Choices Stats:
2025-12-04T09:45:16.0858329Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.0858459Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0858575Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0858737Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0859369Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0859985Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0860650Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0861260Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0861883Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0862489Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0863101Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0863712Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0864344Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0864961Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0865091Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.0865132Z Autotune Choices Stats:
2025-12-04T09:45:16.0865908Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.0866139Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0866307Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0866588Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0867224Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0867856Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0868501Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0869142Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0869776Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0870456Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0871083Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0871719Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0872353Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0873008Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0873152Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.0873228Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0873272Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0873310Z unimplemented []
2025-12-04T09:45:16.0873371Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0873471Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0874057Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0874114Z graph_break []
2025-12-04T09:45:16.0874188Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0874227Z Autotune Choices Stats:
2025-12-04T09:45:16.0874977Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.0875107Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0875223Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0875386Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0876008Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0876633Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0877262Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0877883Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0878490Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0879106Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0879723Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0880348Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0881003Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0881632Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0881776Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.0881817Z Autotune Choices Stats:
2025-12-04T09:45:16.0882582Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.0882802Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0882982Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0883259Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0883899Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0884551Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0885187Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0885842Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0886500Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0887131Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0887774Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0888408Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0889036Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0889759Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0889887Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.0889971Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0890014Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0890051Z unimplemented []
2025-12-04T09:45:16.0890111Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0890212Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0890854Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0890892Z graph_break []
2025-12-04T09:45:16.0890966Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0891007Z Autotune Choices Stats:
2025-12-04T09:45:16.0891757Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.0891902Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0892020Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0892183Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0892816Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0893421Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0894043Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0894668Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0895276Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0895888Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0896511Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0897115Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0897731Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0898349Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0898477Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.0898531Z Autotune Choices Stats:
2025-12-04T09:45:16.0899307Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.0899528Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0899699Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0899990Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0900648Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0901272Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0901904Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0902554Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0903221Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0903858Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0904484Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0905138Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0905777Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0906410Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0906554Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.0906629Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0906671Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0906710Z unimplemented []
2025-12-04T09:45:16.0906769Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0906870Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0907483Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0907534Z graph_break []
2025-12-04T09:45:16.0907609Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0907648Z Autotune Choices Stats:
2025-12-04T09:45:16.0908390Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.0908533Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0908648Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0908810Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0909441Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0910049Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0910695Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0911354Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0911979Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0912582Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0913199Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0913836Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0914449Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0915058Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0915202Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.0915243Z Autotune Choices Stats:
2025-12-04T09:45:16.0916021Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.0916256Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0916422Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0916706Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0917338Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0917982Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0918609Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0919238Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0919889Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0920579Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0921206Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0921840Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0922495Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0923146Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0923275Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.0923349Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0923392Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0923429Z unimplemented []
2025-12-04T09:45:16.0923515Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0923617Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0924198Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0924236Z graph_break []
2025-12-04T09:45:16.0924321Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0924363Z Autotune Choices Stats:
2025-12-04T09:45:16.0925118Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.0925248Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0925369Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0925530Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0926164Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0926777Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0927388Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0927998Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0928626Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0929241Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0929856Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0930505Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0931136Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0931761Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0931894Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.0931937Z Autotune Choices Stats:
2025-12-04T09:45:16.0932694Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.0932933Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0933113Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0933403Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0934049Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0934677Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0935314Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0935949Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0936597Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0937251Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0937892Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0938526Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0939171Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0939815Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0939945Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.0940019Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0940060Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0940099Z unimplemented []
2025-12-04T09:45:16.0940158Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0940258Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0940879Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.0940936Z graph_break []
2025-12-04T09:45:16.0941010Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0941050Z Autotune Choices Stats:
2025-12-04T09:45:16.0941820Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.0941963Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0942079Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0942242Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0942863Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0943482Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0944092Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0944705Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0945317Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0945947Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0946562Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0947172Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0947794Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0948404Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0948537Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.0948576Z Autotune Choices Stats:
2025-12-04T09:45:16.0949352Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.0949580Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0949746Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0950046Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0950731Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0951361Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0951989Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0952637Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0953270Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0953901Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0954565Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0955212Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0955842Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0956480Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0956611Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.0956703Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.0956751Z Traceback (most recent call last):
2025-12-04T09:45:16.0956907Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.0956950Z     self.assertTrue(
2025-12-04T09:45:16.0957058Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.0957108Z     raise self.failureException(msg)
2025-12-04T09:45:16.0957236Z AssertionError: False is not true : Log file /tmp/tmpb10711_l/flex_attention_configs.json was not created
2025-12-04T09:45:16.0957240Z 
2025-12-04T09:45:16.0957316Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.0957485Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.0957487Z 
2025-12-04T09:45:16.0957576Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.0957651Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0957705Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0957745Z unimplemented []
2025-12-04T09:45:16.0957806Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0958385Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.0958496Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0958534Z graph_break []
2025-12-04T09:45:16.0958606Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0959118Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.0959170Z   current_size = base.storage().size()
2025-12-04T09:45:16.0959212Z Autotune Choices Stats:
2025-12-04T09:45:16.0959961Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.0960103Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0960219Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0960380Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0961031Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0961633Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0962240Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0962880Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0963503Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0964105Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0964716Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0965334Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0965959Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0966590Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0966731Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.0966771Z Autotune Choices Stats:
2025-12-04T09:45:16.0967540Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.0967770Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0967941Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0968224Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0968849Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0969487Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0970117Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0970786Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0971432Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0972099Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0972725Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0973353Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0973995Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0974622Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0974755Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.0974828Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0974870Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0974907Z unimplemented []
2025-12-04T09:45:16.0974979Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0975080Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0975665Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0975717Z graph_break []
2025-12-04T09:45:16.0975791Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0975832Z Autotune Choices Stats:
2025-12-04T09:45:16.0976587Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.0976718Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0976836Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0976997Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0977620Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0978223Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0978830Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0979436Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0980065Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0980723Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0981332Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0981942Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0982563Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0983168Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0983299Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.0983338Z Autotune Choices Stats:
2025-12-04T09:45:16.0984103Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.0984350Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0984516Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0984809Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0985441Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0986067Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0986701Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0987347Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0987997Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0988651Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0989283Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0989912Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0990577Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0991215Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0991349Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.0991423Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.0991465Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.0991503Z unimplemented []
2025-12-04T09:45:16.0991563Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.0991663Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.0992236Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.0992288Z graph_break []
2025-12-04T09:45:16.0992361Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.0992402Z Autotune Choices Stats:
2025-12-04T09:45:16.0993162Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.0993304Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.0993420Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.0993581Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.0994204Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0994825Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0995429Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0996050Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0996648Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0997275Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0997890Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.0998497Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0999112Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0999729Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.0999860Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.0999902Z Autotune Choices Stats:
2025-12-04T09:45:16.1000705Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.1000938Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1001106Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1001398Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1002050Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1002682Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1003311Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1003961Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1004596Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1005228Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1005877Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1006513Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1007144Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1007781Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1007909Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.1007982Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1008025Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1008062Z unimplemented []
2025-12-04T09:45:16.1008124Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1008225Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1008797Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1008835Z graph_break []
2025-12-04T09:45:16.1008909Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1008949Z Autotune Choices Stats:
2025-12-04T09:45:16.1009688Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.1009831Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1009959Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1010122Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1010796Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1011402Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1012025Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1012631Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1013237Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1013842Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1014481Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1015096Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1015704Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1016325Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1016456Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.1016497Z Autotune Choices Stats:
2025-12-04T09:45:16.1017260Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.1017480Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1017646Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1017936Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1018588Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1019227Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1019857Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1020542Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1021173Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1021808Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1022438Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1023096Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1023740Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1024367Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1024511Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.1024585Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1024627Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1024666Z unimplemented []
2025-12-04T09:45:16.1024726Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1024826Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1025411Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1025453Z graph_break []
2025-12-04T09:45:16.1025525Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1025570Z Autotune Choices Stats:
2025-12-04T09:45:16.1026321Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.1026460Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1026576Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1026734Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1027361Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1027983Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1028593Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1029210Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1029817Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1030465Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1031074Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1031707Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1032332Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1032938Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1033083Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.1033125Z Autotune Choices Stats:
2025-12-04T09:45:16.1033895Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.1034117Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1034285Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1034561Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1035199Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1035852Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1036492Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1037116Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1037759Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1038396Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1039026Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1039648Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1040303Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1040978Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1041107Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.1041182Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1041226Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1041264Z unimplemented []
2025-12-04T09:45:16.1041327Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1041442Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1042021Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1042060Z graph_break []
2025-12-04T09:45:16.1042136Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1042177Z Autotune Choices Stats:
2025-12-04T09:45:16.1042920Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.1043050Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1043165Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1043324Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1043956Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1044575Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1045196Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1045803Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1046426Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1047044Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1047657Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1048271Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1048904Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1049523Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1049654Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.1049694Z Autotune Choices Stats:
2025-12-04T09:45:16.1050492Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.1050733Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1050900Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1051187Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1051831Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1052453Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1053110Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1053748Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1054378Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1055020Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1055654Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1056287Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1056916Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1057563Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1057706Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.1057781Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1057822Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1057860Z unimplemented []
2025-12-04T09:45:16.1057920Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1058019Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1058593Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1058643Z graph_break []
2025-12-04T09:45:16.1058716Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1058756Z Autotune Choices Stats:
2025-12-04T09:45:16.1059503Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.1059633Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1059749Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1059911Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1060565Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1061193Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1061824Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1062432Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1063036Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1063655Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1064264Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1064880Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1065498Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1066116Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1066259Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.1066300Z Autotune Choices Stats:
2025-12-04T09:45:16.1067056Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.1067293Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1067458Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1067737Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1068366Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1069010Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1069639Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1070290Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1070978Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1071610Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1072260Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1072890Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1073519Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1074171Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1074298Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.1074394Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1074437Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1074474Z unimplemented []
2025-12-04T09:45:16.1074534Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1074637Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1075228Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1075268Z graph_break []
2025-12-04T09:45:16.1075344Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1075383Z Autotune Choices Stats:
2025-12-04T09:45:16.1076133Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.1076275Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1076389Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1076553Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1077170Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1077782Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1078401Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1079029Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1079633Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1080239Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1080895Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1081504Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1082112Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1082740Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1082870Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.1082926Z Autotune Choices Stats:
2025-12-04T09:45:16.1083708Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.1083929Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1084096Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1084398Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1085037Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1085657Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1086285Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1086916Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1087558Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1088200Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1088824Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1089473Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1090104Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1090770Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1090925Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.1091002Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1091046Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1091083Z unimplemented []
2025-12-04T09:45:16.1091144Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1091244Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1091840Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1091893Z graph_break []
2025-12-04T09:45:16.1091967Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1092008Z Autotune Choices Stats:
2025-12-04T09:45:16.1092750Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.1092893Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1093010Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1093174Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1093789Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1094412Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1095027Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1095638Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1096266Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1096867Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1097479Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1098094Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1098703Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1099310Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1099450Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.1099491Z Autotune Choices Stats:
2025-12-04T09:45:16.1100270Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.1100529Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1100696Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1100976Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1101616Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1102260Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1102886Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1103522Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1104166Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1104822Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1105449Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1106093Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1106736Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1107370Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1107498Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.1107573Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1107614Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1107652Z unimplemented []
2025-12-04T09:45:16.1107714Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1107827Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1108403Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1108440Z graph_break []
2025-12-04T09:45:16.1108525Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1108566Z Autotune Choices Stats:
2025-12-04T09:45:16.1109324Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.1109454Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1109569Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1109730Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1110340Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1110992Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1111603Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1112214Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1112861Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1113481Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1114091Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1114696Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1115319Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1115945Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1116074Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.1116115Z Autotune Choices Stats:
2025-12-04T09:45:16.1116880Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.1117114Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1117294Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1117583Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1118212Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1118838Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1119473Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1120097Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1120772Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1121416Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1122081Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1122716Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1123358Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1123999Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1124130Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.1124202Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1124245Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1124283Z unimplemented []
2025-12-04T09:45:16.1124344Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1124444Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1125023Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1125074Z graph_break []
2025-12-04T09:45:16.1125148Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1125190Z Autotune Choices Stats:
2025-12-04T09:45:16.1125948Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.1126078Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1126203Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1126363Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1126984Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1127609Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1128221Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1128827Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1129428Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1130052Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1130721Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1131333Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1131940Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1132565Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1132700Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.1132742Z Autotune Choices Stats:
2025-12-04T09:45:16.1133504Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.1133738Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1133906Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1134185Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1134836Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1135467Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1136099Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1136739Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1137373Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1138026Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1138674Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1139316Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1139938Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1140607Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1140751Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.1140826Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1140867Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1140906Z unimplemented []
2025-12-04T09:45:16.1140966Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1141067Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1141639Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1141678Z graph_break []
2025-12-04T09:45:16.1141752Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1141792Z Autotune Choices Stats:
2025-12-04T09:45:16.1142538Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.1142681Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1142814Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1142976Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1143606Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1144216Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1144840Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1145449Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1146076Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1146684Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1147312Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1147937Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1148549Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1149169Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1149298Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.1149340Z Autotune Choices Stats:
2025-12-04T09:45:16.1150100Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.1150319Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1150516Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1150811Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1151460Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1152098Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1152725Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1153349Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1153990Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1154626Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1155266Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1155921Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1156552Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1157182Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1157325Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.1157400Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1157443Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1157479Z unimplemented []
2025-12-04T09:45:16.1157540Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1157640Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1158218Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1158257Z graph_break []
2025-12-04T09:45:16.1158332Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1158373Z Autotune Choices Stats:
2025-12-04T09:45:16.1159119Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.1159259Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1159374Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1159541Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1160170Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1160831Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1161439Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1162065Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1162668Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1163275Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1163890Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1164535Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1165154Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1165759Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1165901Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.1165943Z Autotune Choices Stats:
2025-12-04T09:45:16.1166700Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.1166919Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1167086Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1167371Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1168010Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1168656Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1169291Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1169914Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1170589Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1171221Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1171853Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1172479Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1173135Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1173774Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1173904Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.1173980Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1174021Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1174062Z unimplemented []
2025-12-04T09:45:16.1174122Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1174226Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1174818Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1174857Z graph_break []
2025-12-04T09:45:16.1174933Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1174973Z Autotune Choices Stats:
2025-12-04T09:45:16.1175721Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.1175848Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1175964Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1176124Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1176750Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1177366Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1177986Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1178592Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1179209Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1179816Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1180465Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1181073Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1181740Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1182373Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1182502Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.1182544Z Autotune Choices Stats:
2025-12-04T09:45:16.1183322Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.1183564Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1183733Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1184019Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1184656Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1185290Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1185933Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1186565Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1187197Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1187839Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1188469Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1189103Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1189729Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1190371Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1190534Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.1190624Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1190667Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1190704Z unimplemented []
2025-12-04T09:45:16.1190764Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1190863Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1191437Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1191489Z graph_break []
2025-12-04T09:45:16.1191563Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1191603Z Autotune Choices Stats:
2025-12-04T09:45:16.1192351Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.1192480Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1192597Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1192758Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1193376Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1193996Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1194614Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1195234Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1198793Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1199476Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1200101Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1200745Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1201354Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1201994Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1202146Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.1202188Z Autotune Choices Stats:
2025-12-04T09:45:16.1202949Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.1203172Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1203364Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1203644Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1204275Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1204904Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1205527Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1206171Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1206808Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1207437Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1208069Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1208693Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1209328Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1209962Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1210093Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.1210182Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1210226Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1210265Z unimplemented []
2025-12-04T09:45:16.1210327Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1210480Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1211170Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1211210Z graph_break []
2025-12-04T09:45:16.1211285Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1211328Z Autotune Choices Stats:
2025-12-04T09:45:16.1212075Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.1212219Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1212336Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1212497Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1213119Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1213723Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1214337Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1214962Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1215565Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1216165Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1216782Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1217388Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1217993Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1218605Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1218733Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.1218775Z Autotune Choices Stats:
2025-12-04T09:45:16.1219554Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.1219773Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1219942Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1220222Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1220908Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1221535Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1222157Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1222777Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1223440Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1224074Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1224695Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1225332Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1225960Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1226586Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1226733Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.1226807Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1226851Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1226888Z unimplemented []
2025-12-04T09:45:16.1226950Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1227050Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1227639Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1227678Z graph_break []
2025-12-04T09:45:16.1227763Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1227803Z Autotune Choices Stats:
2025-12-04T09:45:16.1228549Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.1228677Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1228805Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1228966Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1229582Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1230187Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1230824Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1231453Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1232078Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1232680Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1233285Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1233910Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1234511Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1235117Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1235265Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.1235305Z Autotune Choices Stats:
2025-12-04T09:45:16.1236073Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.1236291Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1236474Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1236749Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1237383Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1238018Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1238637Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1239261Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1239898Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1240591Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1241209Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1241833Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1242478Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1243110Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1243240Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.1243315Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1243356Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1243396Z unimplemented []
2025-12-04T09:45:16.1243456Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1243560Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1244149Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1244188Z graph_break []
2025-12-04T09:45:16.1244260Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1244318Z Autotune Choices Stats:
2025-12-04T09:45:16.1245084Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.1245215Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1245331Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1245489Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1246099Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1246716Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1247320Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1247939Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1248555Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1249179Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1249779Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1250385Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1251070Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1251672Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1251801Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.1251842Z Autotune Choices Stats:
2025-12-04T09:45:16.1252601Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.1252839Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1253032Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1253322Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1253957Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1254579Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1255215Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1255835Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1256464Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1257115Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1257766Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1258393Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1259022Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1259660Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1259788Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.1259861Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1259904Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1259940Z unimplemented []
2025-12-04T09:45:16.1260001Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1260105Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1260724Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1260778Z graph_break []
2025-12-04T09:45:16.1260852Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1260892Z Autotune Choices Stats:
2025-12-04T09:45:16.1261657Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.1261785Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1261914Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1262073Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1262693Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1263310Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1263917Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1264526Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1265130Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1265761Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1266383Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1266988Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1267590Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1268211Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1268341Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.1268381Z Autotune Choices Stats:
2025-12-04T09:45:16.1269147Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.1269364Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1269546Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1269823Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1270622Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1271254Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1271878Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1272533Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1273164Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1273799Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1274452Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1275107Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1275730Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1276353Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1276503Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.1276577Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1276620Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1276658Z unimplemented []
2025-12-04T09:45:16.1276718Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1276818Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1277396Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1277435Z graph_break []
2025-12-04T09:45:16.1277507Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1277549Z Autotune Choices Stats:
2025-12-04T09:45:16.1278284Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.1278426Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1278557Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1278716Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1279344Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1279950Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1280607Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1281214Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1281816Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1282421Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1283065Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1283690Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1284290Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1284896Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1285043Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.1285083Z Autotune Choices Stats:
2025-12-04T09:45:16.1285847Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.1286067Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1286233Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1286510Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1287176Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1287809Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1288430Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1289057Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1289697Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1290324Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1290988Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1291659Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1292302Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1292926Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1293066Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.1293142Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1293183Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1293221Z unimplemented []
2025-12-04T09:45:16.1293281Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1293381Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1293955Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1293993Z graph_break []
2025-12-04T09:45:16.1294066Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1294106Z Autotune Choices Stats:
2025-12-04T09:45:16.1294870Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.1295012Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1295129Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1295291Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1295918Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1296536Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1297145Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1297764Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1298362Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1298962Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1299570Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1300207Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1300881Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1301483Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1301626Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.1301666Z Autotune Choices Stats:
2025-12-04T09:45:16.1302422Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.1302640Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1302812Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1303088Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1303716Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1304376Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1305007Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1305627Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1306276Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1306914Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1307534Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1308154Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1308808Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1309445Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1309575Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.1309648Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1309692Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1309729Z unimplemented []
2025-12-04T09:45:16.1309790Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1309889Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1310518Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1310558Z graph_break []
2025-12-04T09:45:16.1310635Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1310677Z Autotune Choices Stats:
2025-12-04T09:45:16.1311427Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.1311555Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1311671Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1311830Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1312450Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1313100Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1313713Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1314311Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1314929Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1315540Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1316147Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1316749Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1317373Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1317996Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1318125Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.1318166Z Autotune Choices Stats:
2025-12-04T09:45:16.1318927Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.1319156Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1319331Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1319610Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1320251Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1320918Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1321576Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1322220Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1322848Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1323487Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1324116Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1324744Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1325366Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1326012Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1326141Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.1326226Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1326268Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1326307Z unimplemented []
2025-12-04T09:45:16.1326365Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1326472Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1327048Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1327102Z graph_break []
2025-12-04T09:45:16.1327177Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1327217Z Autotune Choices Stats:
2025-12-04T09:45:16.1327960Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.1328087Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1328202Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1328363Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1328979Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1329605Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1330231Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1330891Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1331493Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1332117Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1332735Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1333341Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1333939Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1334576Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1334707Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.1334758Z Autotune Choices Stats:
2025-12-04T09:45:16.1335521Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.1335738Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1335921Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1336202Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1336834Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1337460Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1338082Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1338737Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1339378Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1340018Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1340694Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1341322Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1341950Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1342574Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1342728Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.1342821Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1342865Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1342903Z unimplemented []
2025-12-04T09:45:16.1342966Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1343065Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1343677Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1343716Z graph_break []
2025-12-04T09:45:16.1343788Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1343830Z Autotune Choices Stats:
2025-12-04T09:45:16.1344571Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.1344719Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1344835Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1344999Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1345616Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1346234Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1346858Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1347485Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1348087Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1348693Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1349325Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1349926Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1350565Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1351187Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1351318Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.1351360Z Autotune Choices Stats:
2025-12-04T09:45:16.1352163Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.1352383Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1352551Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1352827Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1353480Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1354104Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1354721Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1355370Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1356031Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1356674Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1357297Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1357963Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1358586Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1359211Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1359360Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.1359437Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1359479Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1359517Z unimplemented []
2025-12-04T09:45:16.1359577Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1359676Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1360268Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1360305Z graph_break []
2025-12-04T09:45:16.1360391Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1360461Z Autotune Choices Stats:
2025-12-04T09:45:16.1361201Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.1361329Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1361473Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1361633Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1362240Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1362849Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1363455Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1364088Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1364719Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1365334Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1365942Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1366563Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1367167Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1367768Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1367915Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.1367954Z Autotune Choices Stats:
2025-12-04T09:45:16.1368744Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.1368963Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1369142Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1369419Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1370052Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1370748Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1371374Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1372000Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1372650Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1373308Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1373933Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1374566Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1375209Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1375835Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1375963Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.1376037Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1376079Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1376116Z unimplemented []
2025-12-04T09:45:16.1376177Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1376292Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1376864Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1376903Z graph_break []
2025-12-04T09:45:16.1376993Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1377033Z Autotune Choices Stats:
2025-12-04T09:45:16.1377789Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.1377923Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1378039Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1378197Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1378810Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1379428Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1380034Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1380680Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1381321Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1381944Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1382551Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1383155Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1383802Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1384408Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1384536Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.1384578Z Autotune Choices Stats:
2025-12-04T09:45:16.1385347Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.1385578Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1385754Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1386048Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1386678Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1387318Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1387959Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1388582Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1389216Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1389866Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1390553Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1391182Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1391808Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1392476Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1392607Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.1392680Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1392722Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1392759Z unimplemented []
2025-12-04T09:45:16.1392819Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1392919Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1393492Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1393548Z graph_break []
2025-12-04T09:45:16.1393622Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1393662Z Autotune Choices Stats:
2025-12-04T09:45:16.1394417Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.1394546Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1394678Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1394839Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1395451Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1396075Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1396680Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1397290Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1397888Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1398525Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1399145Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1399750Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1400357Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1401006Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1401138Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.1401178Z Autotune Choices Stats:
2025-12-04T09:45:16.1401938Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.1402184Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1402349Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1402625Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1403302Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1403929Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1404553Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1405196Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1405826Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1406457Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1407108Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1407747Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1408375Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1409015Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1409143Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.1409219Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1409261Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1409299Z unimplemented []
2025-12-04T09:45:16.1409358Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1409460Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1410034Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1410073Z graph_break []
2025-12-04T09:45:16.1410146Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1410187Z Autotune Choices Stats:
2025-12-04T09:45:16.1410963Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.1411114Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1411245Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1411406Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1412039Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1412648Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1413267Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1413870Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1414475Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1415081Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1415715Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1416335Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1416944Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1417564Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1417691Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.1417732Z Autotune Choices Stats:
2025-12-04T09:45:16.1418495Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.1418714Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1418881Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1419180Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1419820Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1420504Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1421126Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1421768Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1422395Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1423029Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1423652Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1424309Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1424944Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1425577Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1425725Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.1425817Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.1425866Z Traceback (most recent call last):
2025-12-04T09:45:16.1426021Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.1426061Z     self.assertTrue(
2025-12-04T09:45:16.1426170Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.1426220Z     raise self.failureException(msg)
2025-12-04T09:45:16.1426349Z AssertionError: False is not true : Log file /tmp/tmpfzimglfo/flex_attention_configs.json was not created
2025-12-04T09:45:16.1426352Z 
2025-12-04T09:45:16.1426430Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.1426603Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.1426607Z 
2025-12-04T09:45:16.1426698Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.1426772Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1426816Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1426854Z unimplemented []
2025-12-04T09:45:16.1426915Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1427493Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.1427610Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1427649Z graph_break []
2025-12-04T09:45:16.1427722Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1428228Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.1428277Z   current_size = base.storage().size()
2025-12-04T09:45:16.1428318Z Autotune Choices Stats:
2025-12-04T09:45:16.1429086Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.1429217Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1429333Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1429494Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1430119Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1430768Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1431371Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1431970Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1432599Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1433212Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1433818Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1434421Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1435062Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1435666Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1435796Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.1435838Z Autotune Choices Stats:
2025-12-04T09:45:16.1436593Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.1436828Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1437004Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1437301Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1437933Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1438559Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1439198Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1439820Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1440484Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1441161Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1441801Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1442424Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1443045Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1443686Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1443817Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.1443892Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1443934Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1443971Z unimplemented []
2025-12-04T09:45:16.1444032Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1444137Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1444709Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1444763Z graph_break []
2025-12-04T09:45:16.1444837Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1444876Z Autotune Choices Stats:
2025-12-04T09:45:16.1445632Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.1445775Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1445889Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1446049Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1446674Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1447299Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1447899Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1448500Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1449104Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1449732Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1450348Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1450999Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1451628Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1452227Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1452360Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.1452399Z Autotune Choices Stats:
2025-12-04T09:45:16.1453157Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.1453393Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1453560Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1453858Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1454498Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1455133Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1455754Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1456390Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1457017Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1457647Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1458292Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1458938Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1459558Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1460194Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1460324Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.1460398Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1460480Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1460518Z unimplemented []
2025-12-04T09:45:16.1460579Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1460688Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1461261Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1461299Z graph_break []
2025-12-04T09:45:16.1461372Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1461413Z Autotune Choices Stats:
2025-12-04T09:45:16.1462152Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.1462302Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1462429Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1462589Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1463232Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1463833Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1464443Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1465045Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1465650Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1466255Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1466896Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1467509Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1468136Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1468757Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1468884Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.1468927Z Autotune Choices Stats:
2025-12-04T09:45:16.1469687Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.1469907Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1470074Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1470365Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1471086Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1471730Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1472345Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1472962Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1473604Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1474236Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1474859Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1475511Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1476144Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1476766Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1476914Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.1476991Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1477032Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1477070Z unimplemented []
2025-12-04T09:45:16.1477129Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1477232Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1477808Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1477847Z graph_break []
2025-12-04T09:45:16.1477921Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1477960Z Autotune Choices Stats:
2025-12-04T09:45:16.1478710Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.1478856Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1478970Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1479132Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1479757Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1480374Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1481008Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1481633Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1482243Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1482839Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1483448Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1484085Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1484709Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1485305Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1485446Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.1485486Z Autotune Choices Stats:
2025-12-04T09:45:16.1486244Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.1486463Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1486629Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1486903Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1487532Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1488193Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1488827Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1489576Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1490217Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1490886Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1491516Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1492143Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1492805Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1493438Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1493568Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.1493643Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1493686Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1493723Z unimplemented []
2025-12-04T09:45:16.1493783Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1493883Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1494474Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1494511Z graph_break []
2025-12-04T09:45:16.1494584Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1494625Z Autotune Choices Stats:
2025-12-04T09:45:16.1495367Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.1495498Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1495614Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1495774Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1496403Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1497029Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1497650Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1498248Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1498866Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1499462Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1500066Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1500696Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1501336Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1501945Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1502073Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.1502114Z Autotune Choices Stats:
2025-12-04T09:45:16.1502869Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.1503108Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1503275Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1503559Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1504187Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1504810Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1505460Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1506098Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1506725Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1507368Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1507993Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1508619Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1509237Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1509896Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1510038Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.1510114Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1510156Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1510193Z unimplemented []
2025-12-04T09:45:16.1510252Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1510353Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1510969Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1511023Z graph_break []
2025-12-04T09:45:16.1511098Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1511138Z Autotune Choices Stats:
2025-12-04T09:45:16.1511890Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.1512018Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1512133Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1512295Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1512905Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1513519Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1514135Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1514775Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1515374Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1515986Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1516589Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1517192Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1517815Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1521130Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1521291Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.1521331Z Autotune Choices Stats:
2025-12-04T09:45:16.1522087Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.1522306Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1522492Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1522772Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1523405Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1524027Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1524644Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1525305Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1525944Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1526571Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1527206Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1527837Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1528464Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1529104Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1529232Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.1529321Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1529365Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1529402Z unimplemented []
2025-12-04T09:45:16.1529462Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1529562Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1530152Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1530190Z graph_break []
2025-12-04T09:45:16.1530266Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1530308Z Autotune Choices Stats:
2025-12-04T09:45:16.1531081Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.1531226Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1531348Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1531508Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1532121Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1532727Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1533351Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1534007Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1534613Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1535213Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1535833Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1536437Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1537041Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1537662Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1537790Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.1537845Z Autotune Choices Stats:
2025-12-04T09:45:16.1538610Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.1538830Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1538997Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1539277Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1539916Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1540581Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1541203Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1541851Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1542488Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1543118Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1543737Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1544376Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1545007Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1545633Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1545771Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.1545848Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1545889Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1545928Z unimplemented []
2025-12-04T09:45:16.1545987Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1546086Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1546672Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1546710Z graph_break []
2025-12-04T09:45:16.1546796Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1546836Z Autotune Choices Stats:
2025-12-04T09:45:16.1547576Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.1547702Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1547831Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1547998Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1548615Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1549214Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1549815Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1550520Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1551154Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1551754Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1552359Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1552982Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1553582Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1554188Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1554337Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.1554377Z Autotune Choices Stats:
2025-12-04T09:45:16.1555146Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.1555363Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1555539Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1555815Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1556449Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1557085Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1557714Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1558335Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1558971Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1559613Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1560235Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1560912Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1561555Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1562179Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1562309Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.1562383Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1562426Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1562463Z unimplemented []
2025-12-04T09:45:16.1562524Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1562643Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1563212Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1563250Z graph_break []
2025-12-04T09:45:16.1563322Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1563375Z Autotune Choices Stats:
2025-12-04T09:45:16.1564127Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.1564258Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1564376Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1564537Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1565156Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1565776Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1566383Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1566982Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1567607Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1568215Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1568824Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1569430Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1570037Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1570673Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1570802Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.1570844Z Autotune Choices Stats:
2025-12-04T09:45:16.1571607Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.1571840Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1572024Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1572311Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1572946Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1573573Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1574207Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1574831Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1575465Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1576101Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1576740Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1577370Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1577993Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1578624Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1578752Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.1578829Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1578871Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1578909Z unimplemented []
2025-12-04T09:45:16.1578970Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1579071Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1579648Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1579702Z graph_break []
2025-12-04T09:45:16.1579777Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1579817Z Autotune Choices Stats:
2025-12-04T09:45:16.1580600Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.1580728Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1580861Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1581020Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1581630Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1582252Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1582861Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1583462Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1584065Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1584695Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1585316Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1585918Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1586520Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1587138Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1587270Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.1587310Z Autotune Choices Stats:
2025-12-04T09:45:16.1588068Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.1588283Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1588463Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1588742Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1589397Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1590023Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1590685Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1591328Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1591952Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1592582Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1593231Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1593867Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1594496Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1595120Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1595261Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.1595335Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1595378Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1595415Z unimplemented []
2025-12-04T09:45:16.1595476Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1595576Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1596152Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1596189Z graph_break []
2025-12-04T09:45:16.1596263Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1596308Z Autotune Choices Stats:
2025-12-04T09:45:16.1597055Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.1597193Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1597322Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1597485Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1598114Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1598716Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1599340Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1599945Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1600602Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1601219Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1601856Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1602479Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1603094Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1603699Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1603841Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.1603883Z Autotune Choices Stats:
2025-12-04T09:45:16.1604653Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.1604874Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1605042Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1605324Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1605983Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1606620Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1607255Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1607895Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1608534Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1609166Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1609788Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1610495Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1611133Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1611764Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1611907Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.1611984Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1612027Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1612065Z unimplemented []
2025-12-04T09:45:16.1612125Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1612225Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1612802Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1612841Z graph_break []
2025-12-04T09:45:16.1612916Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1612955Z Autotune Choices Stats:
2025-12-04T09:45:16.1613703Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.1613850Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1613967Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1614128Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1614755Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1615373Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1615986Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1616602Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1617221Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1617871Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1618494Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1619135Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1619756Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1620368Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1620575Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.1620616Z Autotune Choices Stats:
2025-12-04T09:45:16.1621372Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.1621590Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1621759Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1622041Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1622680Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1623347Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1623983Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1624610Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1625247Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1625878Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1626512Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1627145Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1627798Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1628437Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1628567Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.1628643Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1628686Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1628723Z unimplemented []
2025-12-04T09:45:16.1628786Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1628884Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1629485Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1629522Z graph_break []
2025-12-04T09:45:16.1629595Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1629636Z Autotune Choices Stats:
2025-12-04T09:45:16.1630378Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.1630555Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1630670Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1630832Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1631457Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1632094Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1632717Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1633330Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1633954Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1634558Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1635170Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1635789Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1636424Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1637034Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1637165Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.1637206Z Autotune Choices Stats:
2025-12-04T09:45:16.1637965Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.1638197Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1638365Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1638647Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1639291Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1639921Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1640625Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1641257Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1641898Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1642553Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1643179Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1643815Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1644439Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1645093Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1645223Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.1645319Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1645363Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1645401Z unimplemented []
2025-12-04T09:45:16.1645461Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1645562Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1646144Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1646193Z graph_break []
2025-12-04T09:45:16.1646269Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1646310Z Autotune Choices Stats:
2025-12-04T09:45:16.1647057Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.1647184Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1647302Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1647462Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1648079Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1648717Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1649334Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1649960Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1650597Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1651217Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1651834Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1652443Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1653056Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1653695Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1653855Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.1653910Z Autotune Choices Stats:
2025-12-04T09:45:16.1654684Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.1654902Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1655083Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1655375Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1656009Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1656638Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1657264Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1657910Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1658555Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1659184Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1659822Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1660487Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1661122Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1661745Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1661893Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.1661967Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1662031Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1662069Z unimplemented []
2025-12-04T09:45:16.1662129Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1662229Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1662829Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1662868Z graph_break []
2025-12-04T09:45:16.1662941Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1662982Z Autotune Choices Stats:
2025-12-04T09:45:16.1663727Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.1663873Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1663988Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1664152Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1664787Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1665395Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1666015Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1666637Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1667251Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1667861Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1668489Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1674099Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1674750Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1675397Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1675537Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.1675580Z Autotune Choices Stats:
2025-12-04T09:45:16.1676374Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.1676609Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1676780Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1677064Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1677719Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1678354Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1678994Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1679621Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1680280Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1680960Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1681595Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1682241Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1682876Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1683508Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1683665Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.1683747Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1683791Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1683831Z unimplemented []
2025-12-04T09:45:16.1683895Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1684004Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1684594Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1684636Z graph_break []
2025-12-04T09:45:16.1684722Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1684763Z Autotune Choices Stats:
2025-12-04T09:45:16.1685525Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.1685654Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1685784Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1685945Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1686572Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1687180Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1687788Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1688405Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1689026Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1689635Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1690251Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1690905Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1691521Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1692144Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1692290Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.1692336Z Autotune Choices Stats:
2025-12-04T09:45:16.1693115Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.1693336Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1693517Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1693801Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1694431Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1695073Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1695707Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1696343Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1696997Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1697647Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1698287Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1698917Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1699579Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1700213Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1700343Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.1700458Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1700504Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1700541Z unimplemented []
2025-12-04T09:45:16.1700603Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1700703Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1701311Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1701348Z graph_break []
2025-12-04T09:45:16.1701422Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1701463Z Autotune Choices Stats:
2025-12-04T09:45:16.1702244Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.1702375Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1702490Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1702651Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1703276Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1703902Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1704513Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1705120Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1705742Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1706372Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1706986Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1707591Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1708220Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1708828Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1708958Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.1708999Z Autotune Choices Stats:
2025-12-04T09:45:16.1709769Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.1710005Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1710193Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1710512Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1711164Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1711793Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1712437Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1713069Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1713705Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1714360Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1715047Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1715677Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1716314Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1716950Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1717079Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.1717157Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1717199Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1717237Z unimplemented []
2025-12-04T09:45:16.1717297Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1717399Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1717976Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1718031Z graph_break []
2025-12-04T09:45:16.1718106Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1718146Z Autotune Choices Stats:
2025-12-04T09:45:16.1718922Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.1719049Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1719183Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1719343Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1719960Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1720606Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1721245Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1721856Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1722458Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1723097Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1723731Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1724339Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1724946Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1725566Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1725696Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.1725738Z Autotune Choices Stats:
2025-12-04T09:45:16.1726509Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.1726735Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1726918Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1727198Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1727863Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1728493Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1729132Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1729781Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1730444Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1731074Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1731723Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1732391Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1733028Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1733656Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1733804Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.1733880Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1733923Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1733960Z unimplemented []
2025-12-04T09:45:16.1734022Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1734121Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1734711Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1734750Z graph_break []
2025-12-04T09:45:16.1734824Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1734871Z Autotune Choices Stats:
2025-12-04T09:45:16.1735609Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.1735755Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1735869Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1736045Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1736679Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1737293Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1737935Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1738546Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1739152Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1739761Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1740445Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1741077Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1741697Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1742316Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1742481Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.1742523Z Autotune Choices Stats:
2025-12-04T09:45:16.1743289Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.1743511Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1743678Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1743955Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1744624Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1745283Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1745921Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1746566Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1747218Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1747852Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1748496Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1749170Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1749825Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1750492Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1750628Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.1750725Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1750768Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1750806Z unimplemented []
2025-12-04T09:45:16.1750866Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1750969Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1751551Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1751590Z graph_break []
2025-12-04T09:45:16.1751664Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1751706Z Autotune Choices Stats:
2025-12-04T09:45:16.1752466Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.1752592Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1752727Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1752888Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1753526Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1754160Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1754773Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1755399Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1756010Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1756636Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1757244Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1757882Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1758499Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1759113Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1759241Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.1759297Z Autotune Choices Stats:
2025-12-04T09:45:16.1760059Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.1760277Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1760486Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1760763Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1761400Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1762070Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1762710Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1763336Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1763982Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1764640Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1765290Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1765923Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1766577Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1767214Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1767346Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.1767419Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1767462Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1767499Z unimplemented []
2025-12-04T09:45:16.1767560Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1767659Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1768257Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1768294Z graph_break []
2025-12-04T09:45:16.1768371Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1768415Z Autotune Choices Stats:
2025-12-04T09:45:16.1769180Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.1769308Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1769422Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1769587Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1770203Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1770863Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1771487Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1772099Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1772716Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1773323Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1773942Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1774564Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1775196Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1775818Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1775948Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.1775990Z Autotune Choices Stats:
2025-12-04T09:45:16.1776758Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.1776989Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1777155Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1777450Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1778091Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1778721Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1779369Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1780013Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1780698Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1781352Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1781985Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1782619Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1783252Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1783916Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1784052Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.1784141Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1784183Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1784221Z unimplemented []
2025-12-04T09:45:16.1784280Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1784380Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1785054Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1785105Z graph_break []
2025-12-04T09:45:16.1785178Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1785227Z Autotune Choices Stats:
2025-12-04T09:45:16.1785974Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.1786100Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1786218Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1786378Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1787000Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1787607Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1788241Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1790221Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1791495Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1792119Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1792736Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1793352Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1793967Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1794592Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1794729Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.1794805Z Autotune Choices Stats:
2025-12-04T09:45:16.1795588Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.1795827Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1795999Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1796284Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1796916Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1797549Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1798184Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1798848Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1799498Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1800143Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1800843Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1801490Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1802133Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1802762Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1802893Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.1802968Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1803026Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1803065Z unimplemented []
2025-12-04T09:45:16.1803125Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1803227Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1803827Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1803878Z graph_break []
2025-12-04T09:45:16.1803952Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1804011Z Autotune Choices Stats:
2025-12-04T09:45:16.1804769Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.1804896Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1805012Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1805175Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1805788Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1806404Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1807012Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1807642Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1808276Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1808891Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1809502Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1810112Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1810769Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1811397Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1811526Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.1811569Z Autotune Choices Stats:
2025-12-04T09:45:16.1812350Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.1812587Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1812762Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1813064Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1813702Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1814331Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1814961Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1815591Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1816234Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1816884Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1817535Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1818165Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1818808Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1819435Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1819564Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.1819642Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1819692Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1819729Z unimplemented []
2025-12-04T09:45:16.1819791Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1819890Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1820521Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1820559Z graph_break []
2025-12-04T09:45:16.1820660Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1820700Z Autotune Choices Stats:
2025-12-04T09:45:16.1821441Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.1821583Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1821700Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1821867Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1822489Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1823101Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1823713Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1824327Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1824964Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1825583Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1826207Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1826822Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1827434Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1828040Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1828170Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.1828213Z Autotune Choices Stats:
2025-12-04T09:45:16.1828990Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.1829213Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1829407Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1829688Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1830338Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1831016Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1831649Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1832280Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1832917Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1833569Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1834224Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1834868Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1835498Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1836116Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1836249Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.1836327Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1836370Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1836409Z unimplemented []
2025-12-04T09:45:16.1836470Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1836572Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1837156Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1837194Z graph_break []
2025-12-04T09:45:16.1837266Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1837319Z Autotune Choices Stats:
2025-12-04T09:45:16.1838074Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.1838217Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1838334Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1838507Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1839133Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1839754Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1840371Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1841035Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1841648Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1842291Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1842917Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1843542Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1844155Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1844761Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1844890Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.1844934Z Autotune Choices Stats:
2025-12-04T09:45:16.1845707Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.1845926Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1846105Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1846382Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1847055Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1847698Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1848324Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1848955Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1849587Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1850225Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1850953Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1851598Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1852244Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1852875Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1853005Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.1853080Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1853123Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1853161Z unimplemented []
2025-12-04T09:45:16.1853225Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1853326Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1853918Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1853955Z graph_break []
2025-12-04T09:45:16.1854030Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1854071Z Autotune Choices Stats:
2025-12-04T09:45:16.1854834Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.1854969Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1855108Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1855270Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1855889Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1856510Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1857124Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1857732Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1858337Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1858955Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1859582Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1860204Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1860858Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1861465Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1861593Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.1861636Z Autotune Choices Stats:
2025-12-04T09:45:16.1862397Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.1862617Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1862787Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1863075Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1863743Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1864392Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1865036Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1865666Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1866302Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1866934Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1867569Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1868210Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1868857Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1869501Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1869630Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.1869708Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1869751Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1869789Z unimplemented []
2025-12-04T09:45:16.1869849Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1869951Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1870571Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1870612Z graph_break []
2025-12-04T09:45:16.1870686Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1870727Z Autotune Choices Stats:
2025-12-04T09:45:16.1871477Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.1871605Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1871747Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1871912Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1872547Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1873169Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1873790Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1874413Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1875021Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1875630Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1876249Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1876870Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1877495Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1878120Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1878248Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.1878291Z Autotune Choices Stats:
2025-12-04T09:45:16.1879055Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.1879275Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1879447Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1879732Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1880382Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1881074Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1881711Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1882353Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1882990Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1883621Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1884257Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1884896Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1885532Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1886166Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1886303Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.1886380Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1886424Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1886463Z unimplemented []
2025-12-04T09:45:16.1886523Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1886630Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1887210Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1887247Z graph_break []
2025-12-04T09:45:16.1887321Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1887363Z Autotune Choices Stats:
2025-12-04T09:45:16.1888117Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.1888247Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1888368Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1888532Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1889155Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1889776Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1890392Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1891057Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1891665Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1892280Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1892901Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1893529Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1894148Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1894770Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1894914Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.1894955Z Autotune Choices Stats:
2025-12-04T09:45:16.1895728Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.1895949Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1896115Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1896396Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1897034Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1897669Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1898304Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1898943Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1899584Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1900223Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1900895Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1901529Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1902174Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1902814Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1902961Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.1903052Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1903096Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1903133Z unimplemented []
2025-12-04T09:45:16.1903196Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1903295Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1903873Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1903913Z graph_break []
2025-12-04T09:45:16.1903990Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1904031Z Autotune Choices Stats:
2025-12-04T09:45:16.1904778Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.1904910Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1905027Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1905191Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1905813Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1906438Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1907069Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1907681Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1908303Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1908918Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1909534Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1910143Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1910803Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1911427Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1911571Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.1911628Z Autotune Choices Stats:
2025-12-04T09:45:16.1912395Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.1912614Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1912782Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1913064Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1913701Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1914332Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1914980Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1915619Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1916260Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1916900Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1917537Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1918164Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1918798Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1919439Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1919587Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.1919681Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.1919728Z Traceback (most recent call last):
2025-12-04T09:45:16.1919884Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.1919923Z     self.assertTrue(
2025-12-04T09:45:16.1920033Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.1920093Z     raise self.failureException(msg)
2025-12-04T09:45:16.1920222Z AssertionError: False is not true : Log file /tmp/tmpqq3rq4tk/flex_attention_configs.json was not created
2025-12-04T09:45:16.1920226Z 
2025-12-04T09:45:16.1920305Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.1920507Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.1920510Z 
2025-12-04T09:45:16.1920602Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.1920677Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1920723Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1920762Z unimplemented []
2025-12-04T09:45:16.1920829Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1921412Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.1921513Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1921550Z graph_break []
2025-12-04T09:45:16.1921626Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1922129Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.1922178Z   current_size = base.storage().size()
2025-12-04T09:45:16.1922218Z Autotune Choices Stats:
2025-12-04T09:45:16.1922968Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.1923100Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1923229Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1923391Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1924017Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1924636Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1925259Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1925864Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1926468Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1927079Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1927711Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1928328Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1928943Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1929558Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1929690Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.1929733Z Autotune Choices Stats:
2025-12-04T09:45:16.1930534Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.1930762Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1930930Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1931213Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1931855Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1932498Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1933139Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1933779Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1934408Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1935038Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1935663Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1936307Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1936944Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1937581Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1937720Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.1937801Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1937843Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1937882Z unimplemented []
2025-12-04T09:45:16.1937941Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1938041Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1938622Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1938662Z graph_break []
2025-12-04T09:45:16.1938739Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1938781Z Autotune Choices Stats:
2025-12-04T09:45:16.1939542Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.1939670Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1939787Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1939948Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1940603Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1941222Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1941843Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1942466Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1943070Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1943680Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1944291Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1944906Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1945525Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1946138Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1946280Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.1946320Z Autotune Choices Stats:
2025-12-04T09:45:16.1947085Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.1947306Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1947474Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1947757Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1948396Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1949043Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1949675Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1950304Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1950997Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1951629Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1952255Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1952886Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1953530Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1954177Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1954319Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.1954417Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1954462Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1954500Z unimplemented []
2025-12-04T09:45:16.1954560Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1954661Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1955245Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1955282Z graph_break []
2025-12-04T09:45:16.1955356Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1955399Z Autotune Choices Stats:
2025-12-04T09:45:16.1956148Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.1956278Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1956394Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1956558Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1957180Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1957807Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1958436Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1959060Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1959665Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1960276Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1960922Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1961530Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1962156Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1962767Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1962910Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.1962965Z Autotune Choices Stats:
2025-12-04T09:45:16.1963730Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.1963952Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1964119Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1964399Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1965037Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1965667Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1966305Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1966950Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1967594Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1968235Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1968868Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1969498Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1970132Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1970813Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1970943Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.1971046Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1971088Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1971128Z unimplemented []
2025-12-04T09:45:16.1971190Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1971297Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1971872Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.1971923Z graph_break []
2025-12-04T09:45:16.1971997Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1972037Z Autotune Choices Stats:
2025-12-04T09:45:16.1972786Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.1972914Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1973043Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1973201Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1973811Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1974410Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1975025Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1975664Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1976277Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1976888Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1977499Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1978105Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1978709Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1979324Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1979463Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.1979514Z Autotune Choices Stats:
2025-12-04T09:45:16.1980279Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.1980558Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1980729Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1981008Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1981646Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1982294Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1982925Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1983566Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1984207Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1984858Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1985506Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1986136Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1986765Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1987392Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1987524Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.1987611Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.1987654Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.1987691Z unimplemented []
2025-12-04T09:45:16.1987753Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.1987852Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.1988444Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.1988481Z graph_break []
2025-12-04T09:45:16.1988566Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.1988607Z Autotune Choices Stats:
2025-12-04T09:45:16.1989348Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.1989484Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1989599Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1989760Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1990378Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1991012Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1991636Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1992280Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1992895Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1993513Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.1994125Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1994741Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1995360Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1995971Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1996102Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.1996142Z Autotune Choices Stats:
2025-12-04T09:45:16.1996936Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.1997169Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.1997347Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.1997625Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.1998256Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1998882Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.1999508Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2000132Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2000815Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2001477Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2002114Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2002758Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2003393Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2004026Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2004155Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.2004237Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2004278Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2004317Z unimplemented []
2025-12-04T09:45:16.2004377Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2004480Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2005075Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2005127Z graph_break []
2025-12-04T09:45:16.2005210Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2005252Z Autotune Choices Stats:
2025-12-04T09:45:16.2006001Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.2006143Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2006260Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2006420Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2007047Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2007657Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2008266Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2008872Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2009505Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2010122Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2010790Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2011403Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2012014Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2012625Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2012757Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.2012802Z Autotune Choices Stats:
2025-12-04T09:45:16.2013578Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.2013798Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2013990Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2014267Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2014915Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2015545Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2016173Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2016798Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2017436Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2018090Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2018727Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2019366Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2019995Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2020749Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2020883Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.2020958Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2021000Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2021037Z unimplemented []
2025-12-04T09:45:16.2021101Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2021206Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2021783Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2021819Z graph_break []
2025-12-04T09:45:16.2021893Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2021954Z Autotune Choices Stats:
2025-12-04T09:45:16.2022716Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.2022862Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2022977Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2023151Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2023775Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2024376Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2024986Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2025598Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2026205Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2026844Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2027466Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2028085Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2028691Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2029304Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2029435Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.2029475Z Autotune Choices Stats:
2025-12-04T09:45:16.2030246Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.2030510Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2030699Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2030989Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2031635Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2032273Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2032901Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2033530Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2034164Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2034794Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2035444Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2036086Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2036725Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2037352Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2037484Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.2037560Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2037601Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2037640Z unimplemented []
2025-12-04T09:45:16.2037699Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2037799Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2038383Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2038422Z graph_break []
2025-12-04T09:45:16.2038495Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2038536Z Autotune Choices Stats:
2025-12-04T09:45:16.2039293Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.2039428Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2039567Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2039728Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2040349Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2041018Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2041632Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2042243Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2042853Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2043492Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2044113Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2044733Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2045360Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2045973Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2046101Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.2046144Z Autotune Choices Stats:
2025-12-04T09:45:16.2046905Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.2047123Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2047293Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2047574Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2048230Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2048870Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2049508Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2050136Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2050818Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2051447Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2052100Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2052744Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2053383Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2054016Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2054144Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.2054219Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2054263Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2054301Z unimplemented []
2025-12-04T09:45:16.2054361Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2054462Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2055043Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2055080Z graph_break []
2025-12-04T09:45:16.2055154Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2055196Z Autotune Choices Stats:
2025-12-04T09:45:16.2055943Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.2056073Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2056201Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2056365Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2057004Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2057617Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2058227Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2058835Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2059445Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2060058Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2060725Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2061344Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2061962Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2062580Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2062710Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.2062750Z Autotune Choices Stats:
2025-12-04T09:45:16.2063515Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.2063738Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2063907Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2064194Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2064832Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2065471Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2066111Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2066748Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2067378Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2068013Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2068638Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2069287Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2069928Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2070607Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2070763Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.2070837Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2070881Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2070919Z unimplemented []
2025-12-04T09:45:16.2070979Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2071079Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2071651Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2071689Z graph_break []
2025-12-04T09:45:16.2071765Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2071806Z Autotune Choices Stats:
2025-12-04T09:45:16.2072555Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.2072686Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2072803Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2072962Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2073595Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2074216Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2074834Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2075444Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2076051Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2076656Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2077263Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2077873Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2078493Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2079107Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2079245Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.2079287Z Autotune Choices Stats:
2025-12-04T09:45:16.2080050Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.2080270Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2080478Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2080759Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2081399Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2082049Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2082685Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2083325Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2083966Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2084606Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2085239Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2085871Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2086517Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2087153Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2087296Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.2087381Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2087424Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2087465Z unimplemented []
2025-12-04T09:45:16.2087526Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2087627Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2088201Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2088238Z graph_break []
2025-12-04T09:45:16.2088312Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2088354Z Autotune Choices Stats:
2025-12-04T09:45:16.2089099Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.2089229Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2089343Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2089504Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2090120Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2090794Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2091421Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2092047Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2092668Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2093278Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2093891Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2094497Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2095115Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2095734Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2095873Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.2095923Z Autotune Choices Stats:
2025-12-04T09:45:16.2096679Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.2096904Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2097073Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2097357Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2097992Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2098622Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2099266Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2099909Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2100591Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2101237Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2101866Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2102500Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2103125Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2103775Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2103908Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.2104018Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2104066Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2104104Z unimplemented []
2025-12-04T09:45:16.2104164Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2104264Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2104841Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2104889Z graph_break []
2025-12-04T09:45:16.2104964Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2105004Z Autotune Choices Stats:
2025-12-04T09:45:16.2105761Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.2105889Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2106006Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2106170Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2106788Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2107387Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2108009Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2108628Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2109245Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2109853Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2110504Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2111112Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2111727Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2112359Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2112490Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.2112558Z Autotune Choices Stats:
2025-12-04T09:45:16.2113320Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.2113553Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2113727Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2114006Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2114634Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2115256Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2115886Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2116522Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2117162Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2117811Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2118446Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2119080Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2119711Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2120338Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2120499Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.2120581Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2120658Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2120699Z unimplemented []
2025-12-04T09:45:16.2120760Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2120860Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2121462Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2121515Z graph_break []
2025-12-04T09:45:16.2121595Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2121652Z Autotune Choices Stats:
2025-12-04T09:45:16.2122402Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.2122529Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2122644Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2122807Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2123417Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2124026Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2124638Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2125257Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2125887Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2126502Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2127111Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2127724Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2128344Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2128956Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2129085Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.2129126Z Autotune Choices Stats:
2025-12-04T09:45:16.2129905Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.2130134Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2130302Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2130640Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2131271Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2131896Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2132524Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2133158Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2133801Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2134443Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2135097Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2135727Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2136357Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2136984Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2137114Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.2137188Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2137235Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2137278Z unimplemented []
2025-12-04T09:45:16.2137341Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2137439Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2138028Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2138067Z graph_break []
2025-12-04T09:45:16.2138165Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2138211Z Autotune Choices Stats:
2025-12-04T09:45:16.2138955Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.2139094Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2139210Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2139375Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2139999Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2140637Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2141248Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2141861Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2142493Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2143110Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2143734Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2144344Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2144954Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2145558Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2145693Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.2145739Z Autotune Choices Stats:
2025-12-04T09:45:16.2146518Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.2146739Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2146924Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2147204Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2147853Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2148483Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2149111Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2149733Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2150368Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2151042Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2151698Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2152338Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2152974Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2153603Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2153731Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.2153805Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2153848Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2153886Z unimplemented []
2025-12-04T09:45:16.2153945Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2154051Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2154630Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2154669Z graph_break []
2025-12-04T09:45:16.2154742Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2154783Z Autotune Choices Stats:
2025-12-04T09:45:16.2155559Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.2155697Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2155812Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2155989Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2156606Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2157214Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2157820Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2158428Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2159044Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2159670Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2160290Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2160939Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2161562Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2162171Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2162300Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.2162341Z Autotune Choices Stats:
2025-12-04T09:45:16.2163108Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.2163327Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2163508Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2163788Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2164444Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2165082Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2165706Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2166334Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2166973Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2167600Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2168255Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2168898Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2169535Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2170175Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2170304Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.2170380Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2170457Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2170499Z unimplemented []
2025-12-04T09:45:16.2170560Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2170660Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2171238Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2171275Z graph_break []
2025-12-04T09:45:16.2171351Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2174047Z Autotune Choices Stats:
2025-12-04T09:45:16.2174833Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.2174965Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2175101Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2175279Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2175900Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2176525Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2177133Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2177741Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2178343Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2178952Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2179597Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2180207Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2180873Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2181483Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2181616Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.2181658Z Autotune Choices Stats:
2025-12-04T09:45:16.2182428Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.2182658Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2182828Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2183110Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2183779Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2184418Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2185057Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2185686Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2186317Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2186948Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2187578Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2188238Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2188873Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2189506Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2189637Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.2189715Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2189759Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2189800Z unimplemented []
2025-12-04T09:45:16.2189861Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2189961Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2190590Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2190628Z graph_break []
2025-12-04T09:45:16.2190702Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2190744Z Autotune Choices Stats:
2025-12-04T09:45:16.2191489Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.2191618Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2191734Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2191915Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2192541Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2193154Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2193771Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2194371Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2194982Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2195590Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2196212Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2196835Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2197454Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2198067Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2198196Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.2198238Z Autotune Choices Stats:
2025-12-04T09:45:16.2198999Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.2199218Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2199386Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2199666Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2200302Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2201026Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2201662Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2202316Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2202951Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2203581Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2204214Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2204869Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2205506Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2206139Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2206279Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.2206355Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2206399Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2206436Z unimplemented []
2025-12-04T09:45:16.2206498Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2206599Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2207176Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2207213Z graph_break []
2025-12-04T09:45:16.2207287Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2207328Z Autotune Choices Stats:
2025-12-04T09:45:16.2208075Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.2208205Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2208320Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2208482Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2209102Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2209717Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2210335Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2210981Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2211585Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2212196Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2212809Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2213440Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2214061Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2214682Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2214824Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.2214865Z Autotune Choices Stats:
2025-12-04T09:45:16.2215622Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.2215843Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2216011Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2216290Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2216921Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2217554Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2218182Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2218827Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2219469Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2220100Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2220766Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2221397Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2222038Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2222674Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2222812Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.2222886Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2222938Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2222979Z unimplemented []
2025-12-04T09:45:16.2223039Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2223139Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2223711Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2223749Z graph_break []
2025-12-04T09:45:16.2223821Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2223863Z Autotune Choices Stats:
2025-12-04T09:45:16.2224612Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.2224740Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2224856Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2225017Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2225632Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2226252Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2226867Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2227483Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2228103Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2228711Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2229317Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2229927Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2230584Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2231208Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2231348Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.2231388Z Autotune Choices Stats:
2025-12-04T09:45:16.2232165Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.2232385Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2232553Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2232834Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2233461Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2234092Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2234742Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2235384Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2236016Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2236655Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2237279Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2237909Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2238537Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2239177Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2239306Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.2239389Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2239442Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2239479Z unimplemented []
2025-12-04T09:45:16.2239541Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2239640Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2240217Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2240263Z graph_break []
2025-12-04T09:45:16.2240336Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2240378Z Autotune Choices Stats:
2025-12-04T09:45:16.2241152Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.2241284Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2241399Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2241562Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2242190Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2242802Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2243418Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2244031Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2244647Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2245263Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2245878Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2246491Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2247097Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2247709Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2247840Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.2247889Z Autotune Choices Stats:
2025-12-04T09:45:16.2248661Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.2248890Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2249056Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2249336Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2249973Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2250631Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2251252Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2251901Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2252546Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2253187Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2253825Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2254461Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2255090Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2255717Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2255846Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.2255919Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2255963Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2256012Z unimplemented []
2025-12-04T09:45:16.2256074Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2256173Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2256756Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2256808Z graph_break []
2025-12-04T09:45:16.2256881Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2256932Z Autotune Choices Stats:
2025-12-04T09:45:16.2257675Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.2257805Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2257920Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2258083Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2258700Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2259306Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2259914Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2260568Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2261199Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2261818Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2262428Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2263028Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2263636Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2264242Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2264373Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.2264413Z Autotune Choices Stats:
2025-12-04T09:45:16.2265193Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.2265422Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2265589Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2265886Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2266525Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2267153Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2267784Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2268415Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2269061Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2269701Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2270334Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2271088Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2271720Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2272348Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2272477Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.2272551Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2272595Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2272635Z unimplemented []
2025-12-04T09:45:16.2272695Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2272795Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2273400Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2273438Z graph_break []
2025-12-04T09:45:16.2273511Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2273582Z Autotune Choices Stats:
2025-12-04T09:45:16.2274326Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.2274467Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2274581Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2274744Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2275364Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2275973Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2276598Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2277197Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2277808Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2278438Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2279064Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2279680Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2280281Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2280944Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2281076Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.2281116Z Autotune Choices Stats:
2025-12-04T09:45:16.2281904Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.2282124Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2282305Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2282604Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2283235Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2283868Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2284501Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2285130Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2285760Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2286401Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2287051Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2287691Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2288321Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2288953Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2289089Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.2289161Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2289205Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2289243Z unimplemented []
2025-12-04T09:45:16.2289303Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2289404Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2289985Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2290023Z graph_break []
2025-12-04T09:45:16.2290095Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2290137Z Autotune Choices Stats:
2025-12-04T09:45:16.2290950Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.2291090Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2291206Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2291378Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2292002Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2292617Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2293225Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2293831Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2294442Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2295077Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2295697Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2296316Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2296929Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2297537Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2297666Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.2297707Z Autotune Choices Stats:
2025-12-04T09:45:16.2298473Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.2298694Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2298869Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2299150Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2299811Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2300481Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2301109Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2301735Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2302368Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2303003Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2303663Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2304308Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2304952Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2305581Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2305710Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.2305786Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2305828Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2305868Z unimplemented []
2025-12-04T09:45:16.2305929Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2306029Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2306604Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2306641Z graph_break []
2025-12-04T09:45:16.2306715Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2306756Z Autotune Choices Stats:
2025-12-04T09:45:16.2307521Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.2307649Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2307774Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2307946Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2308564Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2309182Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2309790Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2310397Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2311040Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2311659Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2312291Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2312905Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2313522Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2314131Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2314262Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.2314304Z Autotune Choices Stats:
2025-12-04T09:45:16.2315073Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.2315291Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2315459Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2315736Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2316392Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2317030Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2317665Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2318293Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2318924Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2319553Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2320181Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2320882Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2321525Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2322166Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2322295Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.2322368Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2322414Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2322450Z unimplemented []
2025-12-04T09:45:16.2322512Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2322611Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2323195Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2323233Z graph_break []
2025-12-04T09:45:16.2323305Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2323346Z Autotune Choices Stats:
2025-12-04T09:45:16.2324095Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.2324223Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2324338Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2324511Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2325134Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2325744Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2326365Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2326973Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2327584Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2328194Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2328814Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2329425Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2330042Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2330696Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2330826Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.2330869Z Autotune Choices Stats:
2025-12-04T09:45:16.2331641Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.2331861Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2332028Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2332310Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2332948Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2333600Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2334240Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2334884Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2335515Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2336149Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2336779Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2337422Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2338062Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2338699Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2338839Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.2338915Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2338957Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2338996Z unimplemented []
2025-12-04T09:45:16.2339056Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2339156Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2339728Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2339766Z graph_break []
2025-12-04T09:45:16.2339842Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2339882Z Autotune Choices Stats:
2025-12-04T09:45:16.2340663Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.2340791Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2340907Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2341067Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2341703Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2342324Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2342944Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2343564Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2344177Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2344796Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2345413Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2346037Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2346653Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2347282Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2347428Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.2347469Z Autotune Choices Stats:
2025-12-04T09:45:16.2348234Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.2348452Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2348622Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2348901Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2349541Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2350173Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2350859Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2351500Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2352145Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2352778Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2353414Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2354046Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2354686Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2355325Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2355464Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.2355548Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2355592Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2355629Z unimplemented []
2025-12-04T09:45:16.2355691Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2355791Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2356363Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2356401Z graph_break []
2025-12-04T09:45:16.2356474Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2356516Z Autotune Choices Stats:
2025-12-04T09:45:16.2357281Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.2357411Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2357527Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2357689Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2358317Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2358939Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2359561Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2360173Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2360835Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2361445Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2362056Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2362667Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2363301Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2363927Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2364068Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.2364125Z Autotune Choices Stats:
2025-12-04T09:45:16.2364892Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.2365112Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2365279Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2365556Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2366196Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2366830Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2367469Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2368102Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2368744Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2369479Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2370108Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2370769Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2371401Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2372043Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2372172Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.2372275Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2372317Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2372359Z unimplemented []
2025-12-04T09:45:16.2372420Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2372521Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2373104Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2373155Z graph_break []
2025-12-04T09:45:16.2373229Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2373269Z Autotune Choices Stats:
2025-12-04T09:45:16.2374014Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.2374142Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2374261Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2374423Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2375047Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2375657Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2376281Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2376910Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2377519Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2378127Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2378740Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2379348Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2379960Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2380627Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2380757Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.2380823Z Autotune Choices Stats:
2025-12-04T09:45:16.2381596Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.2381826Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2381995Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2382275Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2382909Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2383538Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2384158Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2384790Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2385439Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2386081Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2386723Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2387353Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2387983Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2388612Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2388742Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.2388825Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2388868Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2388905Z unimplemented []
2025-12-04T09:45:16.2388970Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2389070Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2389664Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2389702Z graph_break []
2025-12-04T09:45:16.2389787Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2389828Z Autotune Choices Stats:
2025-12-04T09:45:16.2390594Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.2390724Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2390839Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2391002Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2391628Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2392236Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2392845Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2393495Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2394113Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2394739Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2395349Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2395962Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2396572Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2397191Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2397321Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.2397362Z Autotune Choices Stats:
2025-12-04T09:45:16.2398140Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.2398369Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2398546Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2398833Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2399463Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2400092Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2400751Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2401378Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2402019Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2402668Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2403313Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2403940Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2404577Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2405205Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2405334Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.2405413Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2405454Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2405493Z unimplemented []
2025-12-04T09:45:16.2405553Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2405654Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2406247Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2406295Z graph_break []
2025-12-04T09:45:16.2406376Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2406418Z Autotune Choices Stats:
2025-12-04T09:45:16.2407165Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:16.2407302Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2407419Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2407580Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2408199Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2408813Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2409424Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2410030Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2410700Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2411330Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2411952Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2412559Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2413171Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2413781Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2413914Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:16.2413955Z Autotune Choices Stats:
2025-12-04T09:45:16.2414744Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:16.2414963Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2415151Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2415430Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2416083Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2416712Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2417345Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2417973Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2418606Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2419264Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2419897Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2420579Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2421213Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2421843Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2421972Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:16.2422065Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.2422114Z Traceback (most recent call last):
2025-12-04T09:45:16.2422271Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.2422311Z     self.assertTrue(
2025-12-04T09:45:16.2422417Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.2422468Z     raise self.failureException(msg)
2025-12-04T09:45:16.2422596Z AssertionError: False is not true : Log file /tmp/tmpa59m29km/flex_attention_configs.json was not created
2025-12-04T09:45:16.2422599Z 
2025-12-04T09:45:16.2422675Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.2422842Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.2422846Z 
2025-12-04T09:45:16.2422961Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.2423037Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2423081Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2423119Z unimplemented []
2025-12-04T09:45:16.2423180Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2423780Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.2423891Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2423942Z graph_break []
2025-12-04T09:45:16.2424016Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2424512Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.2424561Z   current_size = base.storage().size()
2025-12-04T09:45:16.2424603Z Autotune Choices Stats:
2025-12-04T09:45:16.2425363Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.2425492Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2425612Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2425774Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2426393Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2427005Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2427635Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2428247Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2428861Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2429466Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2430078Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2430702Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2431310Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2431924Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2432077Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.2432119Z Autotune Choices Stats:
2025-12-04T09:45:16.2432877Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.2433113Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2433280Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2433560Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2434193Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2434820Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2435449Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2436078Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2436724Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2437366Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2437990Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2438620Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2439252Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2439876Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2440006Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.2440093Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2440136Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2440173Z unimplemented []
2025-12-04T09:45:16.2440233Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2440334Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2440977Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2441015Z graph_break []
2025-12-04T09:45:16.2441104Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2441144Z Autotune Choices Stats:
2025-12-04T09:45:16.2441894Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.2442025Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2442141Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2442304Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2442921Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2443530Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2444133Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2444777Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2445393Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2446007Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2446618Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2447224Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2447831Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2448435Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2448565Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.2448606Z Autotune Choices Stats:
2025-12-04T09:45:16.2449392Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.2449622Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2449799Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2450081Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2450754Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2451380Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2452003Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2452636Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2453280Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2453933Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2454575Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2455205Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2455839Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2456465Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2456595Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.2456672Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2456714Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2456753Z unimplemented []
2025-12-04T09:45:16.2456813Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2456913Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2457515Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2457562Z graph_break []
2025-12-04T09:45:16.2457644Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2457685Z Autotune Choices Stats:
2025-12-04T09:45:16.2458433Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.2458578Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2458696Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2458858Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2459474Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2460082Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2460727Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2461333Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2461973Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2462598Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2463211Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2463825Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2464440Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2465068Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2465197Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.2465240Z Autotune Choices Stats:
2025-12-04T09:45:16.2466010Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.2466232Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2466418Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2466699Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2467347Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2467976Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2468606Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2469237Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2469868Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2470561Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2471205Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2471844Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2472479Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2473106Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2473236Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.2473313Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2473355Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2473393Z unimplemented []
2025-12-04T09:45:16.2473453Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2473554Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2474143Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2474179Z graph_break []
2025-12-04T09:45:16.2474254Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2474303Z Autotune Choices Stats:
2025-12-04T09:45:16.2475059Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.2475203Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2475317Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2475489Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2476103Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2476709Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2477319Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2477922Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2478533Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2479154Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2479776Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2480428Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2481044Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2481644Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2481774Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.2481816Z Autotune Choices Stats:
2025-12-04T09:45:16.2482588Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.2482807Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2482988Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2483281Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2483930Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2484570Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2485203Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2485848Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2486477Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2487110Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2487759Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2488400Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2489034Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2489667Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2489797Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.2489872Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2489916Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2489953Z unimplemented []
2025-12-04T09:45:16.2490014Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2490114Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2490722Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2490761Z graph_break []
2025-12-04T09:45:16.2490834Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2490875Z Autotune Choices Stats:
2025-12-04T09:45:16.2491634Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.2491763Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2491904Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2492066Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2492683Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2493304Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2493914Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2494522Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2495133Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2495760Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2496373Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2496992Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2497613Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2498225Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2498354Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.2498397Z Autotune Choices Stats:
2025-12-04T09:45:16.2499158Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.2499377Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2499545Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2499832Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2500539Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2501177Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2501819Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2502449Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2503089Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2503729Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2504365Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2505012Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2505650Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2506292Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2506421Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.2506498Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2506540Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2506579Z unimplemented []
2025-12-04T09:45:16.2506638Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2506741Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2507331Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2507368Z graph_break []
2025-12-04T09:45:16.2507442Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2507484Z Autotune Choices Stats:
2025-12-04T09:45:16.2508227Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.2508354Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2508489Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2508650Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2509280Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2509889Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2510562Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2511168Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2511778Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2512394Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2513029Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2513653Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2514276Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2514893Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2515022Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.2515065Z Autotune Choices Stats:
2025-12-04T09:45:16.2515831Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.2516051Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2516220Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2516503Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2517156Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2517797Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2518439Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2519076Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2519709Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2520344Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2521014Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2521664Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2522304Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2522949Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2523099Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.2523174Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2523216Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2523254Z unimplemented []
2025-12-04T09:45:16.2523316Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2523415Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2523996Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2524035Z graph_break []
2025-12-04T09:45:16.2524110Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2524151Z Autotune Choices Stats:
2025-12-04T09:45:16.2524908Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.2525036Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2525152Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2525316Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2525946Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2526562Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2527186Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2527802Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2528409Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2529018Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2529620Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2530241Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2530903Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2531518Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2531659Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.2531702Z Autotune Choices Stats:
2025-12-04T09:45:16.2532468Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.2532688Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2532858Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2533138Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2533778Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2534423Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2535060Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2535691Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2536324Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2536957Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2537583Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2538217Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2538860Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2539497Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2539645Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.2539730Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2539773Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2539812Z unimplemented []
2025-12-04T09:45:16.2539873Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2539978Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2540592Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2540631Z graph_break []
2025-12-04T09:45:16.2540705Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2540747Z Autotune Choices Stats:
2025-12-04T09:45:16.2541502Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.2541628Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2541744Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2541904Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2542523Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2543163Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2543787Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2544402Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2545028Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2545633Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2546242Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2546851Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2547468Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2548087Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2548226Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.2548276Z Autotune Choices Stats:
2025-12-04T09:45:16.2549046Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.2549263Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2549430Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2549711Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2550347Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2551011Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2551656Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2552299Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2552947Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2553583Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2554214Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2554844Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2555483Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2556117Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2556248Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.2556351Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2556395Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2556432Z unimplemented []
2025-12-04T09:45:16.2556495Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2556598Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2557185Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2557233Z graph_break []
2025-12-04T09:45:16.2557309Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2557349Z Autotune Choices Stats:
2025-12-04T09:45:16.2558089Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.2558223Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2558341Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2558504Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2559139Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2559750Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2560375Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2561027Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2561653Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2562258Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2562867Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2563480Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2564091Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2564711Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2564847Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.2564909Z Autotune Choices Stats:
2025-12-04T09:45:16.2565671Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.2565900Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2566067Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2566352Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2566994Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2567629Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2568258Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2568902Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2569550Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2570191Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2570858Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2571499Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2572136Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2572769Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2572897Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.2572972Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2573035Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2573078Z unimplemented []
2025-12-04T09:45:16.2573142Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2573244Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2573835Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2573886Z graph_break []
2025-12-04T09:45:16.2573959Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2574014Z Autotune Choices Stats:
2025-12-04T09:45:16.2574758Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.2574888Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2575006Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2575170Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2575786Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2576399Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2577010Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2577636Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2578265Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2578882Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2579494Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2580108Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2580751Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2581357Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2581486Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.2581527Z Autotune Choices Stats:
2025-12-04T09:45:16.2582318Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.2582548Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2582716Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2583009Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2583648Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2584278Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2584906Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2585530Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2586169Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2586821Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2587464Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2588101Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2588743Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2589370Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2589498Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.2589574Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2589617Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2589654Z unimplemented []
2025-12-04T09:45:16.2589718Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2589818Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2590431Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2590472Z graph_break []
2025-12-04T09:45:16.2590581Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2590621Z Autotune Choices Stats:
2025-12-04T09:45:16.2591362Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.2591505Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2591619Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2591785Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2592407Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2593020Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2593632Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2594242Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2594878Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2595495Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2596117Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2596721Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2597330Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2597940Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2598071Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.2598112Z Autotune Choices Stats:
2025-12-04T09:45:16.2598877Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.2599100Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2599289Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2599567Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2600219Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2600883Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2601507Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2602136Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2602771Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2603410Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2604066Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2604714Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2605341Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2605981Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2606109Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.2606184Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2606227Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2606272Z unimplemented []
2025-12-04T09:45:16.2606332Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2606433Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2607013Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2607051Z graph_break []
2025-12-04T09:45:16.2607124Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2607166Z Autotune Choices Stats:
2025-12-04T09:45:16.2607943Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.2608080Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2608196Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2608365Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2608983Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2609591Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2610197Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2610849Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2611456Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2612102Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2612721Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2613350Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2613960Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2614568Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2614697Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.2614739Z Autotune Choices Stats:
2025-12-04T09:45:16.2615504Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.2615722Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2615907Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2616188Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2616843Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2617481Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2618106Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2618730Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2619374Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2620007Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2620695Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2621343Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2621990Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2622624Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2622754Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.2622830Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2622873Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2622910Z unimplemented []
2025-12-04T09:45:16.2622971Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2623070Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2623672Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2623713Z graph_break []
2025-12-04T09:45:16.2623790Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2623830Z Autotune Choices Stats:
2025-12-04T09:45:16.2624592Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.2624723Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2624854Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2625029Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2625649Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2626268Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2626889Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2627488Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2628098Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2628705Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2629336Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2629962Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2630636Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2631239Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2631370Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.2631413Z Autotune Choices Stats:
2025-12-04T09:45:16.2632184Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.2632405Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2632572Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2632856Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2633541Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2634180Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2634812Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2635444Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2636080Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2636711Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2637351Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2638005Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2638644Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2639283Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2639413Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.2639488Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2639532Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2639573Z unimplemented []
2025-12-04T09:45:16.2639633Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2639733Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2640311Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2640349Z graph_break []
2025-12-04T09:45:16.2640483Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2640525Z Autotune Choices Stats:
2025-12-04T09:45:16.2641282Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.2641415Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2641531Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2641709Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2642343Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2642958Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2643579Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2644188Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2644802Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2645414Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2646032Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2646648Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2647268Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2647884Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2648013Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.2648055Z Autotune Choices Stats:
2025-12-04T09:45:16.2648820Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.2649040Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2649209Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2649494Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2650143Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2650838Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2651482Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2652127Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2652758Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2653393Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2654019Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2654660Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2655309Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2655944Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2656081Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.2656158Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2656201Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2656239Z unimplemented []
2025-12-04T09:45:16.2656299Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2656400Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2656980Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2657017Z graph_break []
2025-12-04T09:45:16.2657094Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2657135Z Autotune Choices Stats:
2025-12-04T09:45:16.2657881Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.2658008Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2658123Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2658288Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2658914Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2659532Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2660153Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2660815Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2661419Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2662024Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2662640Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2663262Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2663882Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2664501Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2664645Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.2664687Z Autotune Choices Stats:
2025-12-04T09:45:16.2665451Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.2665670Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2665837Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2666123Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2666762Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2667404Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2668046Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2668688Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2669330Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2669962Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2670628Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2671263Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2671908Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2672545Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2672688Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.2672762Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2672817Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2672855Z unimplemented []
2025-12-04T09:45:16.2672916Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2673016Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2673599Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2673637Z graph_break []
2025-12-04T09:45:16.2673710Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2673754Z Autotune Choices Stats:
2025-12-04T09:45:16.2674506Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.2674633Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2674748Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2674911Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2675535Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2676144Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2676760Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2677368Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2677979Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2678591Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2679201Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2679816Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2680479Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2681093Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2681230Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.2682891Z Autotune Choices Stats:
2025-12-04T09:45:16.2683682Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.2683905Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2684080Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2684362Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2685001Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2685636Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2686268Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2686910Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2687551Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2688193Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2688824Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2689458Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2690087Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2690770Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2690901Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.2690994Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2691052Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2691092Z unimplemented []
2025-12-04T09:45:16.2691154Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2691258Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2691836Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2691887Z graph_break []
2025-12-04T09:45:16.2691962Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2692005Z Autotune Choices Stats:
2025-12-04T09:45:16.2692758Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.2692887Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2693006Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2693173Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2693788Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2694398Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2695016Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2695631Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2696248Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2696865Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2697476Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2698096Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2698706Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2699316Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2699449Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.2699499Z Autotune Choices Stats:
2025-12-04T09:45:16.2700276Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.2700530Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2700700Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2700983Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2701622Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2702254Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2702878Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2703521Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2704165Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2704807Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2705448Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2706079Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2706713Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2707341Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2707471Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.2707545Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2707590Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2707637Z unimplemented []
2025-12-04T09:45:16.2707699Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2707801Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2708387Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2708435Z graph_break []
2025-12-04T09:45:16.2708508Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2708561Z Autotune Choices Stats:
2025-12-04T09:45:16.2709311Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.2709439Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2709553Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2709717Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2710333Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2710961Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2711568Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2712186Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2712814Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2713429Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2714042Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2714652Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2715261Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2715873Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2716003Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.2716044Z Autotune Choices Stats:
2025-12-04T09:45:16.2716836Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.2717064Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2717233Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2717526Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2718167Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2718802Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2719425Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2720055Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2720756Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2721397Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2722034Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2722684Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2723314Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2723947Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2724076Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.2724149Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2724193Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2724232Z unimplemented []
2025-12-04T09:45:16.2724291Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2724392Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2724993Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2725032Z graph_break []
2025-12-04T09:45:16.2725105Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2725164Z Autotune Choices Stats:
2025-12-04T09:45:16.2725905Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.2726043Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2726159Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2726324Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2726945Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2727551Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2728158Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2728765Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2729381Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2730004Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2730676Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2731292Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2731912Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2732528Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2732748Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.2732818Z Autotune Choices Stats:
2025-12-04T09:45:16.2733629Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.2733871Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2734075Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2734370Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2735016Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2735657Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2736382Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2737015Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2737647Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2738287Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2738965Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2739610Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2740245Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2740941Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2741069Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.2741143Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2741187Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2741240Z unimplemented []
2025-12-04T09:45:16.2741303Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2741405Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2741982Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2742020Z graph_break []
2025-12-04T09:45:16.2742097Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2742139Z Autotune Choices Stats:
2025-12-04T09:45:16.2742938Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.2743083Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2743197Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2743375Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2743995Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2744623Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2745230Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2745838Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2746463Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2747097Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2747720Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2748338Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2748948Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2749554Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2749684Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.2749724Z Autotune Choices Stats:
2025-12-04T09:45:16.2750560Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.2750780Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2750958Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2751238Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2751903Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2752550Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2753175Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2753806Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2754459Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2755095Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2755739Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2756391Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2757025Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2757655Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2757784Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.2757857Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2757901Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2757939Z unimplemented []
2025-12-04T09:45:16.2757998Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2758099Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2758679Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2758717Z graph_break []
2025-12-04T09:45:16.2758790Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2758831Z Autotune Choices Stats:
2025-12-04T09:45:16.2759601Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.2759730Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2759845Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2760027Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2760690Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2761305Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2761911Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2762523Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2763135Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2763746Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2764394Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2765016Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2765640Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2766254Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2766383Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.2766424Z Autotune Choices Stats:
2025-12-04T09:45:16.2767210Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.2767428Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2767595Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2767880Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2768533Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2769183Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2769822Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2770491Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2771122Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2771760Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2772380Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2773037Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2773679Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2774335Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2774465Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.2774538Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2774582Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2774620Z unimplemented []
2025-12-04T09:45:16.2774682Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2774782Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2775358Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2775397Z graph_break []
2025-12-04T09:45:16.2775468Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2775509Z Autotune Choices Stats:
2025-12-04T09:45:16.2776256Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.2776385Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2776499Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2776681Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2777310Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2777922Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2778531Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2779137Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2779746Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2780353Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2781008Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2781666Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2782287Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2782909Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2783039Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.2783080Z Autotune Choices Stats:
2025-12-04T09:45:16.2783855Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.2784076Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2784243Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2784526Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2785166Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2785830Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2786489Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2787135Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2787767Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2788401Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2789039Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2789671Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2790355Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2791043Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2791212Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.2791287Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2791330Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2791369Z unimplemented []
2025-12-04T09:45:16.2791429Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2791530Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2792114Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2792153Z graph_break []
2025-12-04T09:45:16.2792228Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2792269Z Autotune Choices Stats:
2025-12-04T09:45:16.2793045Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.2793184Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2793309Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2793475Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2794149Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2794789Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2795419Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2796035Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2796642Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2797253Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2797887Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2798515Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2799131Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2799763Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2799908Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.2799951Z Autotune Choices Stats:
2025-12-04T09:45:16.2800759Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.2800978Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2801146Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2801431Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2802079Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2802728Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2803366Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2804004Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2804654Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2805290Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2805918Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2806550Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2807194Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2807829Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2807969Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.2808046Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2808099Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2808138Z unimplemented []
2025-12-04T09:45:16.2808200Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2808302Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2808890Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2808929Z graph_break []
2025-12-04T09:45:16.2809003Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2809045Z Autotune Choices Stats:
2025-12-04T09:45:16.2809800Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.2809933Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2810049Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2810213Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2810883Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2811512Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2812131Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2812753Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2813375Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2813983Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2814588Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2815205Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2815826Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2816458Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2816597Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.2816637Z Autotune Choices Stats:
2025-12-04T09:45:16.2817415Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.2817634Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2817799Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2818080Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2818727Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2819358Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2819993Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2820687Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2821333Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2821991Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2822622Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2823262Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2823891Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2824532Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2824661Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.2824746Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2824818Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2824856Z unimplemented []
2025-12-04T09:45:16.2824916Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2825018Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2825600Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2825649Z graph_break []
2025-12-04T09:45:16.2825722Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2825765Z Autotune Choices Stats:
2025-12-04T09:45:16.2826511Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.2826638Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2826757Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2826923Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2827542Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2828155Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2828773Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2829392Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2830006Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2830672Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2831286Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2831901Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2832522Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2833154Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2833283Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.2833337Z Autotune Choices Stats:
2025-12-04T09:45:16.2834114Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.2834344Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2834514Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2834797Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2835442Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2836076Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2836710Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2837347Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2837991Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2838634Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2839279Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2839918Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2840584Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2841214Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2841344Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.2841418Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2841478Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2841516Z unimplemented []
2025-12-04T09:45:16.2841577Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2841677Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2842273Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2842324Z graph_break []
2025-12-04T09:45:16.2842398Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2842451Z Autotune Choices Stats:
2025-12-04T09:45:16.2843279Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.2843409Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2843523Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2843689Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2844301Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2844925Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2845536Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2846162Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2847735Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2848956Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2849586Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2850224Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2850878Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2851485Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2851619Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.2851663Z Autotune Choices Stats:
2025-12-04T09:45:16.2852443Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.2852682Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2852887Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2853181Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2853830Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2854464Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2855094Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2855733Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2856374Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2857009Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2857674Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2858323Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2858954Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2859593Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2859721Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.2859799Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2859843Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2859881Z unimplemented []
2025-12-04T09:45:16.2859942Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2860042Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2860670Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2860710Z graph_break []
2025-12-04T09:45:16.2860794Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2860834Z Autotune Choices Stats:
2025-12-04T09:45:16.2861607Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.2861745Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2861861Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2862034Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2862653Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2863266Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2863887Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2864497Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2865108Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2865738Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2866372Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2866982Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2867611Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2868229Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2868359Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.2868401Z Autotune Choices Stats:
2025-12-04T09:45:16.2869165Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.2869384Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2869569Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2869851Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2870555Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2871206Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2871838Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2872468Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2873099Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2873729Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2874389Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2875037Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2875667Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2876299Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2876428Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.2876502Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2876547Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2876584Z unimplemented []
2025-12-04T09:45:16.2876647Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2876747Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2877349Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2877385Z graph_break []
2025-12-04T09:45:16.2877461Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2877502Z Autotune Choices Stats:
2025-12-04T09:45:16.2878261Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.2878399Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2878525Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2878696Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2879333Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2879941Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2880583Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2881199Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2881815Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2882428Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2883075Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2883706Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2884311Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2884921Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2885051Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.2885093Z Autotune Choices Stats:
2025-12-04T09:45:16.2885858Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.2886078Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2886247Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2886531Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2887195Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2887842Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2888470Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2889101Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2889726Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2890350Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2891013Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2891666Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2892321Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2892950Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2893079Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.2893155Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2893198Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2893236Z unimplemented []
2025-12-04T09:45:16.2893295Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2893397Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2893976Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2894017Z graph_break []
2025-12-04T09:45:16.2894090Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2894132Z Autotune Choices Stats:
2025-12-04T09:45:16.2894881Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.2895009Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2895138Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2895301Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2895934Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2896572Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2897184Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2897793Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2898394Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2899007Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2899620Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2900257Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2900927Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2901535Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2901664Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.2901706Z Autotune Choices Stats:
2025-12-04T09:45:16.2902474Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.2902692Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2902861Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2903144Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2903785Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2904442Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2905094Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2905723Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2906354Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2906984Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2907624Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2908253Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2908909Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2909555Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2909685Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.2909759Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2909803Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2909841Z unimplemented []
2025-12-04T09:45:16.2909901Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2910002Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2910627Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2910664Z graph_break []
2025-12-04T09:45:16.2910741Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2910780Z Autotune Choices Stats:
2025-12-04T09:45:16.2911520Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:16.2911657Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2911772Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2911936Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2912561Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2913218Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2913855Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2914463Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2915075Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2915684Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2916297Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2916912Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2917540Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2918168Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2918299Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:16.2918341Z Autotune Choices Stats:
2025-12-04T09:45:16.2919099Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:16.2919319Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2919486Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2919764Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2920399Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2921086Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2921745Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2922397Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2923032Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2923664Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2924292Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2924927Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2925562Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2926213Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2926351Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:16.2926436Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2926478Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2926517Z unimplemented []
2025-12-04T09:45:16.2926577Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2926678Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2927259Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2927298Z graph_break []
2025-12-04T09:45:16.2927371Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2927412Z Autotune Choices Stats:
2025-12-04T09:45:16.2928163Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.2928291Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2928407Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2928569Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2929187Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2929790Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2930470Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2931102Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2931712Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2932323Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2932943Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2933555Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2934165Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2934790Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2934934Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:16.2934984Z Autotune Choices Stats:
2025-12-04T09:45:16.2935755Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.2935975Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2936142Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2936426Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2937063Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2937684Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2938306Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2938955Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2939608Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2940237Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2940880Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2941513Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2942142Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2942775Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2942916Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:16.2943042Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.2943091Z Traceback (most recent call last):
2025-12-04T09:45:16.2943250Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.2943290Z     self.assertTrue(
2025-12-04T09:45:16.2943398Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.2943447Z     raise self.failureException(msg)
2025-12-04T09:45:16.2943591Z AssertionError: False is not true : Log file /tmp/tmp8_w3cgif/flex_attention_configs.json was not created
2025-12-04T09:45:16.2943595Z 
2025-12-04T09:45:16.2943671Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.2943838Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.2943842Z 
2025-12-04T09:45:16.2943931Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.2944008Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2944051Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2944089Z unimplemented []
2025-12-04T09:45:16.2944150Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2944730Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.2944831Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2944869Z graph_break []
2025-12-04T09:45:16.2944943Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2945441Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.2945492Z   current_size = base.storage().size()
2025-12-04T09:45:16.2945532Z Autotune Choices Stats:
2025-12-04T09:45:16.2946292Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.2946423Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2946550Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2946714Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2947346Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2947966Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2948577Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2949183Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2949792Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2950393Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2951019Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2951656Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2952286Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2952884Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2953019Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.2953061Z Autotune Choices Stats:
2025-12-04T09:45:16.2953817Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.2954040Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2954208Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2954487Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2955127Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2955778Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2956425Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2957053Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2957683Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2958322Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2958953Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2959590Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2960235Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2960920Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2961053Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.2961130Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2961173Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2961210Z unimplemented []
2025-12-04T09:45:16.2961270Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2961369Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2961956Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2961993Z graph_break []
2025-12-04T09:45:16.2962068Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2962108Z Autotune Choices Stats:
2025-12-04T09:45:16.2962861Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.2962990Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2963107Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2963269Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2963896Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2964533Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2965160Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2965765Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2966378Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2966984Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2967596Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2968223Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2968853Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2969470Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2969602Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.2969645Z Autotune Choices Stats:
2025-12-04T09:45:16.2970440Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.2970662Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2970831Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2971112Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2971763Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2972392Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2973043Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2973698Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2974325Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2974956Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2975581Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2976210Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2976833Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2977488Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2977626Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.2977712Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2977754Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2977793Z unimplemented []
2025-12-04T09:45:16.2977854Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2977957Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2978534Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.2978573Z graph_break []
2025-12-04T09:45:16.2978646Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2978687Z Autotune Choices Stats:
2025-12-04T09:45:16.2979435Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.2979562Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2979678Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2979841Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2980509Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2981114Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2981759Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2982381Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2982987Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2983603Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2984213Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2984824Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2985434Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2986064Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2986205Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.2986247Z Autotune Choices Stats:
2025-12-04T09:45:16.2987028Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.2987249Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2987416Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2987698Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2988334Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2988963Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2989596Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2990245Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2990928Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2991560Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2992198Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2992826Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2993459Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2994091Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2994238Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.2994335Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.2994380Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.2994417Z unimplemented []
2025-12-04T09:45:16.2994478Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.2994577Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.2995164Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.2995202Z graph_break []
2025-12-04T09:45:16.2995276Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.2995317Z Autotune Choices Stats:
2025-12-04T09:45:16.2996070Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.2996199Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.2996313Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.2996475Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.2997092Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2997699Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2998316Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.2998947Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.2999574Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3000178Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3000829Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3001437Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3002046Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3002652Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3002809Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.3002888Z Autotune Choices Stats:
2025-12-04T09:45:16.3003669Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.3003890Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3004058Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3004335Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3004969Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3005599Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3006227Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3006852Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3007502Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3008151Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3008776Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3009407Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3010043Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3010717Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3010846Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.3010938Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3010981Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3011019Z unimplemented []
2025-12-04T09:45:16.3011079Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3011179Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3011774Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3011827Z graph_break []
2025-12-04T09:45:16.3011901Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3011954Z Autotune Choices Stats:
2025-12-04T09:45:16.3012694Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.3012821Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3012938Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3013098Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3013721Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3014338Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3014949Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3015566Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3016193Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3016809Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3017422Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3018031Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3018644Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3019250Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3019381Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.3019432Z Autotune Choices Stats:
2025-12-04T09:45:16.3020208Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.3020469Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3020653Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3020939Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3021580Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3022215Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3022842Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3023471Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3024105Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3024763Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3025419Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3026050Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3026681Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3027316Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3027445Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.3027520Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3027563Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3027599Z unimplemented []
2025-12-04T09:45:16.3027660Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3027759Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3028346Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3028392Z graph_break []
2025-12-04T09:45:16.3028476Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3028527Z Autotune Choices Stats:
2025-12-04T09:45:16.3029279Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.3029408Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3029522Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3029683Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3030301Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3030954Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3031564Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3032173Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3032813Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3033444Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3034056Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3034657Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3035268Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3035872Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3036002Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.3036044Z Autotune Choices Stats:
2025-12-04T09:45:16.3036804Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.3037033Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3037219Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3037503Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3038153Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3038777Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3039407Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3040036Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3040706Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3041360Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3042036Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3042665Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3043291Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3043923Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3044051Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.3044128Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3044171Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3044210Z unimplemented []
2025-12-04T09:45:16.3044270Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3044370Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3044951Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3044989Z graph_break []
2025-12-04T09:45:16.3045072Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3045114Z Autotune Choices Stats:
2025-12-04T09:45:16.3045869Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.3046007Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3046134Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3046294Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3046906Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3047514Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3048123Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3048735Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3049335Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3049961Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3050647Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3051254Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3051860Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3052465Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3052596Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.3052638Z Autotune Choices Stats:
2025-12-04T09:45:16.3053411Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.3053631Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3053810Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3054089Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3054752Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3055384Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3056010Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3056637Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3057269Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3057902Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3058565Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3059212Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3059838Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3060507Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3060639Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.3060713Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3060759Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3060796Z unimplemented []
2025-12-04T09:45:16.3060858Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3060959Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3061543Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3061581Z graph_break []
2025-12-04T09:45:16.3061656Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3061696Z Autotune Choices Stats:
2025-12-04T09:45:16.3062443Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.3062597Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3062723Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3062899Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3063530Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3064136Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3064743Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3065357Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3065966Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3066567Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3067202Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3067829Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3068436Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3069048Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3069177Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.3069219Z Autotune Choices Stats:
2025-12-04T09:45:16.3069986Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.3070207Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3070374Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3070705Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3071373Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3072029Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3072654Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3073282Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3073913Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3074545Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3075176Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3075831Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3076483Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3077114Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3077245Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.3077322Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3077365Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3077404Z unimplemented []
2025-12-04T09:45:16.3077464Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3077566Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3078152Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3078193Z graph_break []
2025-12-04T09:45:16.3078267Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3078308Z Autotune Choices Stats:
2025-12-04T09:45:16.3079069Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.3079199Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3079325Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3079486Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3080110Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3080778Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3081391Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3082001Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3082608Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3083217Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3083829Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3084476Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3085113Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3085718Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3085847Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.3085889Z Autotune Choices Stats:
2025-12-04T09:45:16.3086649Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.3086867Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3087035Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3087315Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3087947Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3088595Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3089244Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3089866Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3090533Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3091167Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3091797Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3092434Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3093087Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3093733Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3093863Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.3093938Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3093982Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3094020Z unimplemented []
2025-12-04T09:45:16.3094081Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3094180Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3094766Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3094805Z graph_break []
2025-12-04T09:45:16.3094880Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3094920Z Autotune Choices Stats:
2025-12-04T09:45:16.3095666Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.3095794Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3095909Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3096072Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3096705Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3097332Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3097957Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3098562Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3099173Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3099780Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3100393Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3101034Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3101673Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3102313Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3102443Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.3102483Z Autotune Choices Stats:
2025-12-04T09:45:16.3103251Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.3103470Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3103637Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3103920Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3104552Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3105171Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3105815Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3106461Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3107098Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3107728Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3108348Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3108982Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3109614Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3110259Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3110399Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.3110516Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3110576Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3110616Z unimplemented []
2025-12-04T09:45:16.3110675Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3110776Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3111354Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3111393Z graph_break []
2025-12-04T09:45:16.3111467Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3111507Z Autotune Choices Stats:
2025-12-04T09:45:16.3112251Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.3112381Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3112497Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3112659Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3113278Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3113890Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3114513Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3115136Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3115743Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3116348Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3116958Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3117569Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3118182Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3118810Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3118948Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.3118990Z Autotune Choices Stats:
2025-12-04T09:45:16.3119753Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.3119972Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3120144Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3120466Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3121102Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3121739Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3122372Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3123030Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3123688Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3124319Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3124945Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3125571Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3126202Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3126837Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3126975Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.3127072Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3127126Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3127165Z unimplemented []
2025-12-04T09:45:16.3127226Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3127327Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3127926Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3127963Z graph_break []
2025-12-04T09:45:16.3128038Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3128080Z Autotune Choices Stats:
2025-12-04T09:45:16.3128834Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.3128963Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3129078Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3129244Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3129859Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3130508Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3131121Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3131756Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3132387Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3132996Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3133616Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3134225Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3134832Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3135440Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3135579Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.3135619Z Autotune Choices Stats:
2025-12-04T09:45:16.3136411Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.3136633Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3136799Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3137081Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3137721Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3138355Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3138989Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3139612Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3140262Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3140949Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3141577Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3142212Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3142846Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3143473Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3143602Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.3143688Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3143732Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3143769Z unimplemented []
2025-12-04T09:45:16.3143833Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3143933Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3144531Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3144583Z graph_break []
2025-12-04T09:45:16.3144657Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3144699Z Autotune Choices Stats:
2025-12-04T09:45:16.3145461Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.3145590Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3145706Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3145870Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3146486Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3147093Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3147705Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3148319Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3148946Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3149565Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3150181Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3150830Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3151434Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3152046Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3152175Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.3152242Z Autotune Choices Stats:
2025-12-04T09:45:16.3153022Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.3153260Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3153439Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3153719Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3154368Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3154998Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3155632Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3156263Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3156895Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3157545Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3158191Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3158831Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3159461Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3160093Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3160222Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.3160299Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3160342Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3160382Z unimplemented []
2025-12-04T09:45:16.3160479Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3160582Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3161163Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3161217Z graph_break []
2025-12-04T09:45:16.3161292Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3161358Z Autotune Choices Stats:
2025-12-04T09:45:16.3162118Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.3162246Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3162363Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3162528Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3163160Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3163785Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3164388Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3164989Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3165607Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3166243Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3166862Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3167471Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3168079Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3168689Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3168820Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.3168859Z Autotune Choices Stats:
2025-12-04T09:45:16.3169623Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.3169853Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3170037Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3170324Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3171024Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3171657Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3172284Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3172915Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3173546Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3174180Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3174831Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3175484Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3176115Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3176738Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3176868Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.3176942Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3176986Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3177024Z unimplemented []
2025-12-04T09:45:16.3177085Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3177184Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3177763Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3177802Z graph_break []
2025-12-04T09:45:16.3177874Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3177924Z Autotune Choices Stats:
2025-12-04T09:45:16.3178682Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.3178818Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3178934Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3179108Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3179724Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3180331Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3180976Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3181590Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3182193Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3182823Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3183462Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3184077Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3184692Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3189238Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3189390Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.3189446Z Autotune Choices Stats:
2025-12-04T09:45:16.3190216Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.3190478Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3190698Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3190982Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3191646Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3192304Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3192930Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3193559Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3194191Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3194819Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3195463Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3196112Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3196747Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3197366Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3197499Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.3197580Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3197627Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3197667Z unimplemented []
2025-12-04T09:45:16.3197731Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3197834Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3198420Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3198460Z graph_break []
2025-12-04T09:45:16.3198536Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3198578Z Autotune Choices Stats:
2025-12-04T09:45:16.3199332Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.3199473Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3199602Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3199778Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3200455Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3201066Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3201675Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3202280Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3202890Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3203500Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3204155Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3204803Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3205418Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3206022Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3206156Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.3206196Z Autotune Choices Stats:
2025-12-04T09:45:16.3206956Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.3207175Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3207347Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3207628Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3208284Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3208938Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3209570Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3210217Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3210885Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3211510Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3212141Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3212807Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3213463Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3214089Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3214220Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.3214295Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3214339Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3214378Z unimplemented []
2025-12-04T09:45:16.3214441Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3214541Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3215121Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3215160Z graph_break []
2025-12-04T09:45:16.3215232Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3215274Z Autotune Choices Stats:
2025-12-04T09:45:16.3216019Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.3216148Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3216263Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3216435Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3217056Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3217682Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3218288Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3218895Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3219505Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3220111Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3220758Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3221392Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3222024Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3222630Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3222760Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.3222800Z Autotune Choices Stats:
2025-12-04T09:45:16.3223564Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.3223784Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3223951Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3224233Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3224885Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3225540Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3226189Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3226816Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3227451Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3228076Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3228705Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3229345Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3229995Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3230701Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3230830Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.3230905Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3230949Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3230988Z unimplemented []
2025-12-04T09:45:16.3231048Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3231151Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3231733Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3231773Z graph_break []
2025-12-04T09:45:16.3231845Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3231888Z Autotune Choices Stats:
2025-12-04T09:45:16.3232636Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.3232763Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3232881Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3233043Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3233671Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3234307Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3234937Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3235542Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3236148Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3236748Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3237351Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3237955Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3238589Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3239213Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3239343Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.3239384Z Autotune Choices Stats:
2025-12-04T09:45:16.3240156Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.3240374Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3240578Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3240860Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3241497Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3242130Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3242802Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3243463Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3244097Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3244736Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3245362Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3245987Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3246620Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3247268Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3247408Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.3247482Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3247526Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3247574Z unimplemented []
2025-12-04T09:45:16.3247636Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3247735Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3248307Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3248344Z graph_break []
2025-12-04T09:45:16.3248417Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3248457Z Autotune Choices Stats:
2025-12-04T09:45:16.3249202Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.3249329Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3249443Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3249607Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3250226Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3250863Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3251507Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3252137Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3252743Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3253357Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3253972Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3254582Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3255189Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3255823Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3255963Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.3256004Z Autotune Choices Stats:
2025-12-04T09:45:16.3256780Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.3257002Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3257169Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3257452Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3258091Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3258727Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3259352Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3259997Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3260688Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3261310Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3261936Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3262583Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3263344Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3263972Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3264115Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.3264191Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3264257Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3264298Z unimplemented []
2025-12-04T09:45:16.3264358Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3264460Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3265055Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3265096Z graph_break []
2025-12-04T09:45:16.3265169Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3265211Z Autotune Choices Stats:
2025-12-04T09:45:16.3265959Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.3266086Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3266200Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3266364Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3266985Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3267599Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3268204Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3268837Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3269471Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3270072Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3270786Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3271399Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3272011Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3272623Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3272766Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.3272808Z Autotune Choices Stats:
2025-12-04T09:45:16.3273609Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.3273830Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3273998Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3274280Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3274920Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3275555Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3276181Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3276801Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3277458Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3278106Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3278734Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3279365Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3280000Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3280668Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3280801Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.3280874Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3280939Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3280977Z unimplemented []
2025-12-04T09:45:16.3281039Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3281138Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3281752Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3281803Z graph_break []
2025-12-04T09:45:16.3281876Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3281917Z Autotune Choices Stats:
2025-12-04T09:45:16.3282675Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.3282804Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3282917Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3283083Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3283701Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3284319Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3284939Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3285554Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3286179Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3286823Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3287435Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3288045Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3288653Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3289262Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3289391Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.3289441Z Autotune Choices Stats:
2025-12-04T09:45:16.3290211Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.3290469Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3290648Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3290929Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3291571Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3292202Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3292836Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3293472Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3294106Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3294773Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3295418Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3296055Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3296699Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3297316Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3297446Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.3297521Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3297565Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3297603Z unimplemented []
2025-12-04T09:45:16.3297662Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3297763Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3298342Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3298394Z graph_break []
2025-12-04T09:45:16.3298467Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3298508Z Autotune Choices Stats:
2025-12-04T09:45:16.3299286Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.3299415Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3299529Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3299692Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3300314Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3300948Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3301553Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3302159Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3302787Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3303432Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3304056Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3304665Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3305275Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3305883Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3306012Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.3306052Z Autotune Choices Stats:
2025-12-04T09:45:16.3306817Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.3307053Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3307232Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3307518Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3308160Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3308788Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3309413Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3310045Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3310725Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3311361Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3312023Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3312681Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3313314Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3313940Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3314067Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.3314141Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3314186Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3314224Z unimplemented []
2025-12-04T09:45:16.3314285Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3314385Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3314965Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3315003Z graph_break []
2025-12-04T09:45:16.3315077Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3315128Z Autotune Choices Stats:
2025-12-04T09:45:16.3315889Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.3316028Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3316141Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3316313Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3316937Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3317548Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3318159Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3318771Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3319378Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3320014Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3320689Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3321301Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3321915Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3322523Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3322653Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.3322694Z Autotune Choices Stats:
2025-12-04T09:45:16.3323461Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.3323682Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3323849Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3324139Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3324787Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3325436Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3326071Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3326700Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3327338Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3327973Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3328607Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3329267Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3329900Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3330573Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3330703Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.3330775Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3330820Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3330857Z unimplemented []
2025-12-04T09:45:16.3330918Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3331018Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3331595Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3331633Z graph_break []
2025-12-04T09:45:16.3331705Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3331746Z Autotune Choices Stats:
2025-12-04T09:45:16.3332497Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.3332648Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3332777Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3332952Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3333584Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3334197Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3334808Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3335423Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3336028Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3336640Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3337272Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3337906Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3338520Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3339127Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3339258Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.3339299Z Autotune Choices Stats:
2025-12-04T09:45:16.3340067Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.3340287Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3340489Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3340769Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3341433Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3342088Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3342717Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3343344Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3343979Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3344619Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3345253Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3345909Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3346555Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3347185Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3347315Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.3347390Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3347432Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3347471Z unimplemented []
2025-12-04T09:45:16.3347530Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3347630Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3348214Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3348253Z graph_break []
2025-12-04T09:45:16.3348326Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3348366Z Autotune Choices Stats:
2025-12-04T09:45:16.3349117Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.3349245Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3349359Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3349535Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3350167Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3350842Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3351452Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3352071Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3352685Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3353291Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3353902Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3354537Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3355167Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3355770Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3355900Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.3355941Z Autotune Choices Stats:
2025-12-04T09:45:16.3356712Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.3356932Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3357098Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3357378Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3358016Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3358664Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3359314Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3359944Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3360609Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3361240Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3361870Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3362509Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3363167Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3363820Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3363949Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.3364024Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3364068Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3364106Z unimplemented []
2025-12-04T09:45:16.3364168Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3364268Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3364850Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3364888Z graph_break []
2025-12-04T09:45:16.3364961Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3365004Z Autotune Choices Stats:
2025-12-04T09:45:16.3365759Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.3365886Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3366003Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3366165Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3366797Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3367434Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3368065Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3368670Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3369280Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3369882Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3370531Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3371145Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3371776Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3372407Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3372536Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.3372577Z Autotune Choices Stats:
2025-12-04T09:45:16.3373345Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.3373565Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3373733Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3374014Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3374659Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3375291Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3375944Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3376594Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3377225Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3377858Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3378489Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3379121Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3379755Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3380439Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3380586Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.3380661Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3380725Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3380765Z unimplemented []
2025-12-04T09:45:16.3380824Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3380925Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3381497Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3381535Z graph_break []
2025-12-04T09:45:16.3381609Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3381649Z Autotune Choices Stats:
2025-12-04T09:45:16.3382393Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.3382520Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3382635Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3382800Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3383409Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3384019Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3384671Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3385297Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3385909Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3386520Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3387135Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3387746Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3388356Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3388988Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3389128Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.3389169Z Autotune Choices Stats:
2025-12-04T09:45:16.3389948Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.3390166Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3390335Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3390673Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3391313Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3391951Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3392583Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3393243Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3393903Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3394536Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3395164Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3395793Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3396423Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3397049Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3397190Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.3397272Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3397327Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3397364Z unimplemented []
2025-12-04T09:45:16.3397425Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3397525Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3398117Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3398157Z graph_break []
2025-12-04T09:45:16.3398231Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3398272Z Autotune Choices Stats:
2025-12-04T09:45:16.3399011Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.3399140Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3399256Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3399416Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3400027Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3400689Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3401296Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3401935Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3402565Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3403174Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3403783Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3404400Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3405015Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3405622Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3405761Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.3405834Z Autotune Choices Stats:
2025-12-04T09:45:16.3406608Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.3406828Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3406998Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3407279Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3407914Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3408549Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3409179Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3409814Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3410510Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3411164Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3411799Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3412428Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3413060Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3413693Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3413822Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.3413923Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3413966Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3414004Z unimplemented []
2025-12-04T09:45:16.3414065Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3414165Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3414753Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3414809Z graph_break []
2025-12-04T09:45:16.3414885Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3414935Z Autotune Choices Stats:
2025-12-04T09:45:16.3415683Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.3415811Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3415930Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3416092Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3416724Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3417341Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3417948Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3418567Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3419194Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3419814Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3420467Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3421081Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3421692Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3422303Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3422434Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.3422490Z Autotune Choices Stats:
2025-12-04T09:45:16.3423270Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.3423502Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3423684Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3423966Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3424603Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3425231Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3425858Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3426488Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3427124Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3427777Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3428426Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3429060Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3429696Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3430335Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3430515Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.3430592Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3430634Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3430672Z unimplemented []
2025-12-04T09:45:16.3430733Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3430834Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3431437Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3431475Z graph_break []
2025-12-04T09:45:16.3431566Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3431620Z Autotune Choices Stats:
2025-12-04T09:45:16.3432382Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:16.3432511Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3432626Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3432789Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3433405Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3434016Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3434626Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3435232Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3435862Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3436487Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3437100Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3437710Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3438323Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3438923Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3439053Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:16.3439096Z Autotune Choices Stats:
2025-12-04T09:45:16.3439851Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:16.3440083Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3440269Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3440592Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3441255Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3441885Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3442512Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3443142Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3443776Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3444432Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3445080Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3445719Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3446365Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3446999Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3447128Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:16.3447204Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3447245Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3447283Z unimplemented []
2025-12-04T09:45:16.3447343Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3447444Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3448025Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3448082Z graph_break []
2025-12-04T09:45:16.3448157Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3448198Z Autotune Choices Stats:
2025-12-04T09:45:16.3448957Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.3449093Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3449219Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3449378Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3449999Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3450643Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3451252Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3451869Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3452477Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3453110Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3453751Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3454359Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3454970Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3455578Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3455707Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:16.3455750Z Autotune Choices Stats:
2025-12-04T09:45:16.3456525Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.3456836Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3457016Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3457307Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3457963Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3458585Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3459219Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3459849Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3460514Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3461142Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3461792Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3462455Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3463083Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3463711Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3463841Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:16.3463915Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3463957Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3463994Z unimplemented []
2025-12-04T09:45:16.3464056Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3464158Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3464758Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3464795Z graph_break []
2025-12-04T09:45:16.3464869Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3464911Z Autotune Choices Stats:
2025-12-04T09:45:16.3465659Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.3465801Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3465948Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3466109Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3466738Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3467345Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3467960Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3468583Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3469190Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3469797Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3470487Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3471122Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3471733Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3472341Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3472471Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:16.3472512Z Autotune Choices Stats:
2025-12-04T09:45:16.3473288Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.3473508Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3473674Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3473955Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3474621Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3475268Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3475898Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3476527Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3477160Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3477793Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3478424Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3479075Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3479741Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3480374Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3480549Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:16.3480647Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.3480694Z Traceback (most recent call last):
2025-12-04T09:45:16.3480850Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.3480891Z     self.assertTrue(
2025-12-04T09:45:16.3480998Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.3481050Z     raise self.failureException(msg)
2025-12-04T09:45:16.3481181Z AssertionError: False is not true : Log file /tmp/tmpv8xw9256/flex_attention_configs.json was not created
2025-12-04T09:45:16.3481184Z 
2025-12-04T09:45:16.3481260Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.3481427Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.3481431Z 
2025-12-04T09:45:16.3481521Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.3481598Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3481642Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3481679Z unimplemented []
2025-12-04T09:45:16.3481741Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3482334Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.3482453Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3482489Z graph_break []
2025-12-04T09:45:16.3482564Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3483073Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.3483137Z   current_size = base.storage().size()
2025-12-04T09:45:16.3483179Z Autotune Choices Stats:
2025-12-04T09:45:16.3483945Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.3484076Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3484192Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3484354Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3484973Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3485573Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3486183Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3486792Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3487411Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3488039Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3488652Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3489259Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3489869Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3490512Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3490644Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.3490686Z Autotune Choices Stats:
2025-12-04T09:45:16.3491454Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.3491700Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3491879Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3492177Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3492818Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3493442Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3494067Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3494694Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3495329Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3495984Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3496627Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3497263Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3497891Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3498508Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3498639Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.3498715Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3498757Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3498796Z unimplemented []
2025-12-04T09:45:16.3498857Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3498957Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3499544Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3499595Z graph_break []
2025-12-04T09:45:16.3499669Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3499714Z Autotune Choices Stats:
2025-12-04T09:45:16.3500508Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.3500651Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3500780Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3500941Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3501566Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3502169Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3502796Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3503399Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3504014Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3504641Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3505266Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3505879Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3506491Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3507096Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3507225Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.3507267Z Autotune Choices Stats:
2025-12-04T09:45:16.3508027Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.3508247Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3508426Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3508712Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3509359Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3509986Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3510640Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3511270Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3511918Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3512556Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3513215Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3513865Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3514494Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3515123Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3515253Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.3515328Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3515371Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3515408Z unimplemented []
2025-12-04T09:45:16.3515468Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3515567Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3516154Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3516190Z graph_break []
2025-12-04T09:45:16.3516266Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3516307Z Autotune Choices Stats:
2025-12-04T09:45:16.3517052Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.3517192Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3517329Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3517490Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3518118Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3518725Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3519339Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3519948Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3520593Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3521206Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3521836Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3522469Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3523074Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3523682Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3523815Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.3523857Z Autotune Choices Stats:
2025-12-04T09:45:16.3524621Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.3524842Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3525010Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3525293Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3525954Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3526608Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3527237Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3527865Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3528496Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3529132Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3529765Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3530457Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3531113Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3531741Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3531871Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.3531948Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3531990Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3532028Z unimplemented []
2025-12-04T09:45:16.3532088Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3532187Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3532767Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3532805Z graph_break []
2025-12-04T09:45:16.3532877Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3532919Z Autotune Choices Stats:
2025-12-04T09:45:16.3533663Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.3533791Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3533921Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3534081Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3534711Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3535340Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3535950Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3536565Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3537178Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3537786Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3538395Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3539015Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3539646Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3540257Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3540386Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.3540467Z Autotune Choices Stats:
2025-12-04T09:45:16.3541229Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.3541448Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3541616Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3541902Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3542534Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3543262Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3543918Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3544552Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3545183Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3545814Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3546441Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3547073Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3547715Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3548359Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3548488Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.3548565Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3548607Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3548646Z unimplemented []
2025-12-04T09:45:16.3548706Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3548809Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3549394Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3549431Z graph_break []
2025-12-04T09:45:16.3549506Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3549547Z Autotune Choices Stats:
2025-12-04T09:45:16.3550293Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.3550462Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3550577Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3550741Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3551361Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3551985Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3552612Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3553216Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3553823Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3554429Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3555045Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3555661Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3556290Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3556912Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3557042Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.3557084Z Autotune Choices Stats:
2025-12-04T09:45:16.3557861Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.3558081Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3558249Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3558528Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3559166Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3559798Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3560490Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3561140Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3561769Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3562402Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3563027Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3563659Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3564289Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3564952Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3565091Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.3565176Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3565220Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3565257Z unimplemented []
2025-12-04T09:45:16.3565318Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3565417Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3565997Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3566037Z graph_break []
2025-12-04T09:45:16.3566111Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3566152Z Autotune Choices Stats:
2025-12-04T09:45:16.3566899Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.3567028Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3567149Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3567309Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3567930Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3568544Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3569168Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3569793Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3570393Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3571030Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3571632Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3572239Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3572847Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3573481Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3573622Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.3573664Z Autotune Choices Stats:
2025-12-04T09:45:16.3574443Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.3574665Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3574835Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3575115Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3575754Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3576380Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3577009Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3577660Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3578304Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3578939Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3579564Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3580198Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3580865Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3581495Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3581645Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.3581746Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3581788Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3581825Z unimplemented []
2025-12-04T09:45:16.3581885Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3581988Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3582581Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3582619Z graph_break []
2025-12-04T09:45:16.3582695Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3582737Z Autotune Choices Stats:
2025-12-04T09:45:16.3583482Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.3583609Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3583728Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3583893Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3584509Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3585116Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3585727Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3586350Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3586974Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3587588Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3588197Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3588803Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3589415Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3590020Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3590161Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.3590224Z Autotune Choices Stats:
2025-12-04T09:45:16.3591047Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.3591266Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3591438Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3591720Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3592356Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3592984Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3593614Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3594242Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3594897Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3595553Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3596185Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3596816Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3597450Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3598080Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3598209Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.3598294Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3598339Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3598376Z unimplemented []
2025-12-04T09:45:16.3598436Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3598536Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3599135Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3599185Z graph_break []
2025-12-04T09:45:16.3599257Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3599308Z Autotune Choices Stats:
2025-12-04T09:45:16.3600053Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.3600182Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3600297Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3600484Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3601099Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3601699Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3602313Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3602935Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3603582Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3604200Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3604804Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3605414Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3606023Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3606626Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3606757Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.3606810Z Autotune Choices Stats:
2025-12-04T09:45:16.3607581Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.3607822Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3608003Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3608280Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3608926Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3609557Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3610185Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3610874Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3611503Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3612168Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3612821Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3613460Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3614094Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3614727Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3614856Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.3614931Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3614972Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3615011Z unimplemented []
2025-12-04T09:45:16.3615070Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3615171Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3615751Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3615804Z graph_break []
2025-12-04T09:45:16.3615889Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3615941Z Autotune Choices Stats:
2025-12-04T09:45:16.3616700Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.3616832Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3616949Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3617112Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3617736Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3618348Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3618956Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3619567Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3620194Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3620869Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3621488Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3622104Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3622716Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3623323Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3623454Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.3623496Z Autotune Choices Stats:
2025-12-04T09:45:16.3624262Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.3624496Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3624682Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3624974Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3625621Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3626252Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3626885Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3627511Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3628147Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3628780Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3629425Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3630070Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3630741Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3631372Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3631501Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.3631576Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3631622Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3631659Z unimplemented []
2025-12-04T09:45:16.3631719Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3631819Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3632397Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3632442Z graph_break []
2025-12-04T09:45:16.3632532Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3632574Z Autotune Choices Stats:
2025-12-04T09:45:16.3633350Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.3633491Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3633623Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3633787Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3634406Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3635014Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3635622Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3636235Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3636841Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3637466Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3638099Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3638709Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3639321Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3639925Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3640056Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.3640097Z Autotune Choices Stats:
2025-12-04T09:45:16.3640912Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.3641132Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3641321Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3641598Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3642256Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3642895Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3643517Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3644147Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3644783Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3645418Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3646063Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3646709Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3647342Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3647970Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3648098Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.3648174Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3648216Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3648255Z unimplemented []
2025-12-04T09:45:16.3648316Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3648416Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3648995Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3649035Z graph_break []
2025-12-04T09:45:16.3649109Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3649152Z Autotune Choices Stats:
2025-12-04T09:45:16.3649904Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.3650044Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3650171Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3650344Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3651012Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3651625Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3652246Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3652857Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3653464Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3654093Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3654732Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3655367Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3655981Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3656593Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3656723Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.3656768Z Autotune Choices Stats:
2025-12-04T09:45:16.3657534Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.3657755Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3657926Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3658208Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3658873Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3659526Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3660151Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3660813Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3661450Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3662084Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3662712Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3663371Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3664025Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3664659Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3664789Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.3664864Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3664909Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3664945Z unimplemented []
2025-12-04T09:45:16.3665007Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3665108Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3665687Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3665724Z graph_break []
2025-12-04T09:45:16.3665799Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3665839Z Autotune Choices Stats:
2025-12-04T09:45:16.3666590Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.3666723Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3666837Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3667010Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3667639Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3668271Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3668876Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3669484Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3670090Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3670732Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3671340Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3671969Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3672608Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3673221Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3673352Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.3673392Z Autotune Choices Stats:
2025-12-04T09:45:16.3674151Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.3674373Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3674542Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3674823Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3675464Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3676112Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3676753Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3677385Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3678030Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3678661Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3679293Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3679928Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3680609Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3681265Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3681395Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.3681471Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3681516Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3681555Z unimplemented []
2025-12-04T09:45:16.3681615Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3681716Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3682298Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3682338Z graph_break []
2025-12-04T09:45:16.3682411Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3682455Z Autotune Choices Stats:
2025-12-04T09:45:16.3683201Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.3683328Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3683447Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3683609Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3684231Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3684862Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3685488Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3686096Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3686710Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3687330Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3687945Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3688559Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3689188Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3689824Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3689953Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.3689996Z Autotune Choices Stats:
2025-12-04T09:45:16.3690791Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.3691009Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3691178Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3691462Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3692101Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3692730Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3693386Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3694040Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3694675Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3695302Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3695933Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3696562Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3697198Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3699784Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3702786Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.3702866Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3702939Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3702978Z unimplemented []
2025-12-04T09:45:16.3703042Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3703144Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3703730Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3703769Z graph_break []
2025-12-04T09:45:16.3703845Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3703886Z Autotune Choices Stats:
2025-12-04T09:45:16.3704638Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.3704768Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3704884Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3705048Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3705663Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3706270Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3706906Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3707525Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3708125Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3708728Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3709335Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3709939Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3710583Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3711220Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3711363Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.3711404Z Autotune Choices Stats:
2025-12-04T09:45:16.3712178Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.3712399Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3712568Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3712847Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3713473Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3714098Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3714718Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3715364Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3716010Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3716643Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3717270Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3717899Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3718525Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3719148Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3719290Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.3719377Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3719431Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3719470Z unimplemented []
2025-12-04T09:45:16.3719531Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3719632Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3720216Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3720255Z graph_break []
2025-12-04T09:45:16.3720328Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3720371Z Autotune Choices Stats:
2025-12-04T09:45:16.3721140Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.3721268Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3721385Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3721545Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3722159Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3722763Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3723370Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3724013Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3724647Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3725256Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3725860Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3726463Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3727066Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3727673Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3727816Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.3727859Z Autotune Choices Stats:
2025-12-04T09:45:16.3728637Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.3728856Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3729023Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3729298Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3729930Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3730583Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3731206Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3731831Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3732483Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3733125Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3733746Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3734385Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3735004Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3735630Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3735758Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.3735848Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3735892Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3735930Z unimplemented []
2025-12-04T09:45:16.3735989Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3736091Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3736672Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3736718Z graph_break []
2025-12-04T09:45:16.3736792Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3736832Z Autotune Choices Stats:
2025-12-04T09:45:16.3737578Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.3737707Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3737821Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3737985Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3738598Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3739204Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3739815Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3740446Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3741076Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3741705Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3742312Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3742917Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3743522Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3744122Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3744251Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.3744303Z Autotune Choices Stats:
2025-12-04T09:45:16.3745075Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.3745302Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3745479Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3745757Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3746388Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3747017Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3747639Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3748266Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3748903Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3749547Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3750189Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3750849Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3751473Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3752098Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3752228Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.3752301Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3752346Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3752384Z unimplemented []
2025-12-04T09:45:16.3752446Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3752545Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3753122Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3753184Z graph_break []
2025-12-04T09:45:16.3753256Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3753333Z Autotune Choices Stats:
2025-12-04T09:45:16.3754078Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.3754207Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3754324Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3754486Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3755102Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3755702Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3756311Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3756918Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3757531Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3758160Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3758784Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3759401Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3760010Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3760656Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3760784Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.3760829Z Autotune Choices Stats:
2025-12-04T09:45:16.3761591Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.3761823Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3762005Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3762294Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3762940Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3763574Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3764198Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3764828Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3765464Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3766092Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3766731Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3767379Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3768014Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3768644Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3768772Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.3768848Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3768890Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3768929Z unimplemented []
2025-12-04T09:45:16.3768990Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3769091Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3769674Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3769713Z graph_break []
2025-12-04T09:45:16.3769789Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3769838Z Autotune Choices Stats:
2025-12-04T09:45:16.3770638Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.3770780Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3770894Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3771071Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3771692Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3772301Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3772917Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3773539Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3774162Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3774792Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3775423Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3776031Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3776640Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3777240Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3777371Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.3777413Z Autotune Choices Stats:
2025-12-04T09:45:16.3778190Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.3778410Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3778589Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3778871Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3779507Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3780156Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3780841Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3781463Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3782096Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3782727Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3783383Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3784038Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3784670Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3785305Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3785435Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.3785508Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3785554Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3785594Z unimplemented []
2025-12-04T09:45:16.3785655Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3785756Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3786337Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3786376Z graph_break []
2025-12-04T09:45:16.3786448Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3786495Z Autotune Choices Stats:
2025-12-04T09:45:16.3787250Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.3787393Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3787520Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3787691Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3788327Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3788937Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3789552Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3790179Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3790829Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3791435Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3792077Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3792710Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3793322Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3793932Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3794061Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.3794103Z Autotune Choices Stats:
2025-12-04T09:45:16.3794873Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.3795093Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3795261Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3795544Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3796204Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3796853Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3797482Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3798116Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3798747Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3799381Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3800012Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3800724Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3801377Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3802008Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3802138Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.3802216Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3802260Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3802300Z unimplemented []
2025-12-04T09:45:16.3802360Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3802465Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3803048Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3803087Z graph_break []
2025-12-04T09:45:16.3803162Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3803202Z Autotune Choices Stats:
2025-12-04T09:45:16.3803957Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.3804085Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3804202Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3804377Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3805009Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3805653Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3806256Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3806869Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3807482Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3808092Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3808707Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3809338Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3809963Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3810613Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3810743Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.3810784Z Autotune Choices Stats:
2025-12-04T09:45:16.3811556Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.3811774Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3811942Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3812226Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3812855Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3813510Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3814162Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3814790Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3815433Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3816067Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3816700Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3817333Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3817985Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3818633Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3818763Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.3818836Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3818882Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3818919Z unimplemented []
2025-12-04T09:45:16.3818981Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3819081Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3819670Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3819708Z graph_break []
2025-12-04T09:45:16.3819780Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3819823Z Autotune Choices Stats:
2025-12-04T09:45:16.3820607Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.3820735Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3820849Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3821013Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3821636Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3822278Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3822908Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3823518Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3824133Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3824744Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3825373Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3826000Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3826636Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3827257Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3827388Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.3827430Z Autotune Choices Stats:
2025-12-04T09:45:16.3828193Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.3828411Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3828580Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3828863Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3829507Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3830150Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3830836Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3831483Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3832110Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3832742Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3833369Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3834007Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3834636Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3835285Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3835423Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.3835497Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3835540Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3835587Z unimplemented []
2025-12-04T09:45:16.3835648Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3835749Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3836327Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3836366Z graph_break []
2025-12-04T09:45:16.3836439Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3836479Z Autotune Choices Stats:
2025-12-04T09:45:16.3837229Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.3837357Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3837473Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3837634Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3838256Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3838856Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3839492Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3840110Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3840744Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3841356Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3841967Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3842572Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3843181Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3843828Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3843972Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.3844014Z Autotune Choices Stats:
2025-12-04T09:45:16.3844812Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.3845031Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3845200Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3845484Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3846122Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3846754Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3847378Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3848027Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3848679Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3849312Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3849945Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3850609Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3851237Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3851859Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3852008Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.3852081Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3852153Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3852191Z unimplemented []
2025-12-04T09:45:16.3852251Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3852350Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3852957Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3852996Z graph_break []
2025-12-04T09:45:16.3853069Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3853111Z Autotune Choices Stats:
2025-12-04T09:45:16.3853862Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.3853991Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3854105Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3854269Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3854891Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3855505Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3856110Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3856730Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3857356Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3857968Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3858574Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3859183Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3859790Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3860394Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3860571Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.3860612Z Autotune Choices Stats:
2025-12-04T09:45:16.3861424Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.3861644Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3861810Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3862092Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3862734Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3863369Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3864002Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3864635Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3865290Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3865943Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3866570Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3867209Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3867841Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3868469Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3868599Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.3868684Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3868726Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3868765Z unimplemented []
2025-12-04T09:45:16.3868824Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3868926Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3869521Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3869577Z graph_break []
2025-12-04T09:45:16.3869649Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3869691Z Autotune Choices Stats:
2025-12-04T09:45:16.3870497Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.3870626Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3870741Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3870904Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3871528Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3872129Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3872740Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3873347Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3873970Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3874599Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3875214Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3875824Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3876441Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3877050Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3877178Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.3877230Z Autotune Choices Stats:
2025-12-04T09:45:16.3878012Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.3878238Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3878416Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3878694Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3879336Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3879970Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3880648Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3881279Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3881913Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3882566Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3883218Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3883850Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3884483Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3885112Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3885241Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.3885315Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3885359Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3885396Z unimplemented []
2025-12-04T09:45:16.3885457Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3885558Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3886142Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3886189Z graph_break []
2025-12-04T09:45:16.3886264Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3886332Z Autotune Choices Stats:
2025-12-04T09:45:16.3887081Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.3887211Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3887325Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3887489Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3888104Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3888718Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3889327Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3889943Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3890601Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3891240Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3891864Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3892475Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3893097Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3893720Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3893853Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.3893893Z Autotune Choices Stats:
2025-12-04T09:45:16.3894659Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.3894892Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3895067Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3895355Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3896003Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3896632Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3897263Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3897892Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3898516Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3899149Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3899803Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3900481Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3901115Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3901765Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3901897Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.3901976Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3902019Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3902057Z unimplemented []
2025-12-04T09:45:16.3902117Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3902217Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3902801Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3902842Z graph_break []
2025-12-04T09:45:16.3902933Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3902975Z Autotune Choices Stats:
2025-12-04T09:45:16.3903733Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.3903873Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3903988Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3904162Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3904786Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3905396Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3906015Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3906631Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3907238Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3907863Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3908491Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3909101Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3909710Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3910315Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3910473Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.3910516Z Autotune Choices Stats:
2025-12-04T09:45:16.3911297Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.3911515Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3911700Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3911977Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3912628Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3913264Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3913894Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3914521Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3915155Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3915786Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3916438Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3917089Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3917722Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3918367Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3918498Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.3918572Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3918617Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3918655Z unimplemented []
2025-12-04T09:45:16.3918717Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3918816Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3919393Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3919430Z graph_break []
2025-12-04T09:45:16.3919505Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3919547Z Autotune Choices Stats:
2025-12-04T09:45:16.3920301Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.3920477Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3920605Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3920780Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3921412Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3922024Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3922634Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3923247Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3923860Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3924474Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3925108Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3925736Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3926344Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3926966Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3927097Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.3927140Z Autotune Choices Stats:
2025-12-04T09:45:16.3927913Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.3928133Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3928298Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3928577Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3929232Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3929878Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3930533Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3931161Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3931794Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3932433Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3933062Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3933714Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3934376Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3935003Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3935141Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.3935217Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3935261Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3935300Z unimplemented []
2025-12-04T09:45:16.3935360Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3935461Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3936049Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3936088Z graph_break []
2025-12-04T09:45:16.3936162Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3936205Z Autotune Choices Stats:
2025-12-04T09:45:16.3936955Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.3937084Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3937211Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3937370Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3937999Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3938622Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3939232Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3939846Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3940489Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3941098Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3941712Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3942349Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3942985Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3943599Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3943730Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.3943771Z Autotune Choices Stats:
2025-12-04T09:45:16.3944539Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.3944760Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3944929Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3945208Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3945849Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3946495Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3947142Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3947769Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3948407Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3949037Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3949672Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3950305Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3950975Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3951627Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3951755Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.3951831Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3951873Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3951910Z unimplemented []
2025-12-04T09:45:16.3951971Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3952072Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3952653Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3952691Z graph_break []
2025-12-04T09:45:16.3952766Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3952807Z Autotune Choices Stats:
2025-12-04T09:45:16.3953553Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.3953683Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3953798Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3953960Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3954577Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3955212Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3955842Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3956448Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3957056Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3957664Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3958275Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3958888Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3959525Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3960151Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3960281Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.3960324Z Autotune Choices Stats:
2025-12-04T09:45:16.3961133Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.3961352Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3961520Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3961801Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3962439Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3963067Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3963719Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3964373Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3965007Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3965656Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3966285Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3966912Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3967563Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3968216Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3968356Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.3968443Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3968487Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3968526Z unimplemented []
2025-12-04T09:45:16.3968584Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3968685Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3969275Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.3969314Z graph_break []
2025-12-04T09:45:16.3969386Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3969427Z Autotune Choices Stats:
2025-12-04T09:45:16.3970182Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:16.3970312Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3970444Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3970602Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3971220Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3971824Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3972461Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3973093Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3973701Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3974326Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3974934Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3975545Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3976170Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3976797Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3976936Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:16.3976977Z Autotune Choices Stats:
2025-12-04T09:45:16.3977751Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:16.3977972Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3978141Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3978418Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3979062Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3979694Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3980317Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3981008Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3981665Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3982296Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3982926Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3983562Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3984191Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3984820Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3984961Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:16.3985056Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.3985098Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.3985137Z unimplemented []
2025-12-04T09:45:16.3985196Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.3985298Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.3985888Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.3985925Z graph_break []
2025-12-04T09:45:16.3986001Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.3986041Z Autotune Choices Stats:
2025-12-04T09:45:16.3986790Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.3986918Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3987034Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3987195Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3987808Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3988416Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3989035Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3989670Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3990285Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3990920Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.3991536Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3992149Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3992766Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3993376Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3993528Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:16.3993580Z Autotune Choices Stats:
2025-12-04T09:45:16.3994356Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.3994577Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.3994744Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.3995026Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.3995671Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3996305Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3996940Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3997562Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3998210Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3998872Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.3999496Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4000127Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4000790Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4001420Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4001548Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:16.4001635Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4001678Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4001715Z unimplemented []
2025-12-04T09:45:16.4001774Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4001873Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4002459Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4002511Z graph_break []
2025-12-04T09:45:16.4002596Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4002639Z Autotune Choices Stats:
2025-12-04T09:45:16.4003391Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.4003520Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4003638Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4003800Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4004432Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4005046Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4005668Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4006294Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4006913Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4007524Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4009116Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4009743Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4010362Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4011012Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4011165Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:16.4011206Z Autotune Choices Stats:
2025-12-04T09:45:16.4011987Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.4012209Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4012389Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4012670Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4013325Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4013990Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4014620Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4015266Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4015930Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4016577Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4017214Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4017841Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4018489Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4019117Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4019247Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:16.4019323Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4019365Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4019404Z unimplemented []
2025-12-04T09:45:16.4019464Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4019566Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4020156Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4020195Z graph_break []
2025-12-04T09:45:16.4020278Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4020319Z Autotune Choices Stats:
2025-12-04T09:45:16.4021119Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:16.4021248Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4021365Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4021526Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4022159Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4022768Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4023379Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4023988Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4024627Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4025249Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4025861Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4026489Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4027100Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4027710Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4027840Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:16.4027881Z Autotune Choices Stats:
2025-12-04T09:45:16.4028655Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.4028895Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4029063Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4029355Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4029990Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4030662Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4031313Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4031939Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4032585Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4033240Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4033882Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4034514Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4035151Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4035782Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4035913Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:16.4036006Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.4036055Z Traceback (most recent call last):
2025-12-04T09:45:16.4036211Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.4036251Z     self.assertTrue(
2025-12-04T09:45:16.4036358Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.4036408Z     raise self.failureException(msg)
2025-12-04T09:45:16.4036538Z AssertionError: False is not true : Log file /tmp/tmpcnpjpknz/flex_attention_configs.json was not created
2025-12-04T09:45:16.4036542Z 
2025-12-04T09:45:16.4036620Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.4036785Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.4036800Z 
2025-12-04T09:45:16.4036891Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.4036966Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4037010Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4037049Z unimplemented []
2025-12-04T09:45:16.4037113Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4037718Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.4037822Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4037868Z graph_break []
2025-12-04T09:45:16.4037943Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4038442Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.4038490Z   current_size = base.storage().size()
2025-12-04T09:45:16.4038531Z Autotune Choices Stats:
2025-12-04T09:45:16.4039296Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.4039429Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4039546Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4039709Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4040335Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4040999Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4041632Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4042245Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4042849Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4043453Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4044080Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4044704Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4045313Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4045927Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4046070Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.4046114Z Autotune Choices Stats:
2025-12-04T09:45:16.4046898Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.4047121Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4047287Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4047576Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4048214Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4048850Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4049478Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4050100Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4050787Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4051430Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4052052Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4052691Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4053320Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4053952Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4054095Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.4054171Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4054213Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4054253Z unimplemented []
2025-12-04T09:45:16.4054313Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4054417Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4055005Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4055044Z graph_break []
2025-12-04T09:45:16.4055127Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4055168Z Autotune Choices Stats:
2025-12-04T09:45:16.4055914Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.4056053Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4056168Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4056331Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4056944Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4057542Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4058152Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4058769Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4059379Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4059983Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4060624Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4061226Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4061835Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4062442Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4062587Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.4062628Z Autotune Choices Stats:
2025-12-04T09:45:16.4063419Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.4063639Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4063817Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4064095Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4064721Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4065360Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4065987Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4066612Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4067249Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4067892Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4068526Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4069154Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4069792Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4070455Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4070585Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.4070658Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4070701Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4070739Z unimplemented []
2025-12-04T09:45:16.4070801Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4070899Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4071491Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4071529Z graph_break []
2025-12-04T09:45:16.4071614Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4071655Z Autotune Choices Stats:
2025-12-04T09:45:16.4072418Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.4072549Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4072665Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4072826Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4073454Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4074074Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4074683Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4075287Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4075913Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4076535Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4077140Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4077755Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4078362Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4078988Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4079121Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.4079161Z Autotune Choices Stats:
2025-12-04T09:45:16.4079924Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.4080166Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4080341Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4080658Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4081312Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4081941Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4082585Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4083213Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4083844Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4084500Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4085139Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4085765Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4086402Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4087029Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4087157Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.4087233Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4087276Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4087314Z unimplemented []
2025-12-04T09:45:16.4087374Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4087476Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4088048Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4088096Z graph_break []
2025-12-04T09:45:16.4088171Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4088211Z Autotune Choices Stats:
2025-12-04T09:45:16.4088972Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.4089101Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4089225Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4089385Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4089999Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4090671Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4091277Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4091887Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4092494Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4093127Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4093749Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4094357Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4094982Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4095588Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4095720Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.4095760Z Autotune Choices Stats:
2025-12-04T09:45:16.4096526Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.4096746Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4096922Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4097218Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4097864Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4098491Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4099126Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4099759Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4100391Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4101097Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4101746Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4102393Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4103026Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4103660Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4103788Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.4103861Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4103905Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4103943Z unimplemented []
2025-12-04T09:45:16.4104005Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4104105Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4104688Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4104726Z graph_break []
2025-12-04T09:45:16.4104802Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4104843Z Autotune Choices Stats:
2025-12-04T09:45:16.4105594Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.4105733Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4105856Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4106017Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4106643Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4107246Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4107863Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4108461Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4109072Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4109703Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4110331Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4111000Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4111610Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4112229Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4112358Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.4112399Z Autotune Choices Stats:
2025-12-04T09:45:16.4113157Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.4113377Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4113545Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4113825Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4114488Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4115125Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4115751Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4116386Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4117037Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4117700Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4118327Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4118981Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4119618Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4120249Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4120387Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.4120487Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4120528Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4120567Z unimplemented []
2025-12-04T09:45:16.4120627Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4120728Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4121307Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4121344Z graph_break []
2025-12-04T09:45:16.4121420Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4121460Z Autotune Choices Stats:
2025-12-04T09:45:16.4122196Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.4122324Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4122453Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4122614Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4123253Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4123868Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4124478Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4125096Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4125704Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4126313Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4126922Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4127557Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4128172Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4128777Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4128918Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.4128959Z Autotune Choices Stats:
2025-12-04T09:45:16.4129725Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.4129946Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4130114Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4130390Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4131059Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4131711Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4132349Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4132979Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4133618Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4134243Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4134873Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4135511Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4136158Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4136804Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4136934Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.4137008Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4137052Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4137091Z unimplemented []
2025-12-04T09:45:16.4137162Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4137262Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4137838Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4137875Z graph_break []
2025-12-04T09:45:16.4137949Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4137989Z Autotune Choices Stats:
2025-12-04T09:45:16.4138736Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.4138865Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4138980Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4139139Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4139756Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4140380Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4141045Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4141651Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4142269Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4142878Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4143492Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4144103Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4144737Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4145360Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4145492Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.4145533Z Autotune Choices Stats:
2025-12-04T09:45:16.4146285Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.4146517Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4146684Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4146966Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4147604Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4148228Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4148874Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4149510Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4150144Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4150807Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4151433Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4152064Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4152697Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4153351Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4153480Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.4153567Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4153609Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4153648Z unimplemented []
2025-12-04T09:45:16.4153707Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4153808Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4154385Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4154439Z graph_break []
2025-12-04T09:45:16.4154514Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4154555Z Autotune Choices Stats:
2025-12-04T09:45:16.4155304Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.4155432Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4155550Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4155714Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4156330Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4156938Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4157562Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4158180Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4158784Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4159414Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4160030Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4160662Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4161271Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4161902Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4162034Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.4162075Z Autotune Choices Stats:
2025-12-04T09:45:16.4162861Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.4163090Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4163259Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4163539Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4164172Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4164795Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4165424Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4166068Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4166710Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4167343Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4167984Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4168629Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4169260Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4169889Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4170028Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.4170113Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4170157Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4170193Z unimplemented []
2025-12-04T09:45:16.4170256Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4170355Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4170984Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4171023Z graph_break []
2025-12-04T09:45:16.4171096Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4171136Z Autotune Choices Stats:
2025-12-04T09:45:16.4171874Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.4172016Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4172130Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4172291Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4172909Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4173513Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4174134Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4174746Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4175371Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4175973Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4176600Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4177210Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4177833Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4178437Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4178578Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.4178628Z Autotune Choices Stats:
2025-12-04T09:45:16.4179397Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.4179619Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4179785Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4180073Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4180757Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4181385Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4182008Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4182632Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4183287Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4183934Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4184550Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4185195Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4185834Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4186470Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4186600Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.4186686Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4186730Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4186771Z unimplemented []
2025-12-04T09:45:16.4186831Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4186931Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4187513Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4187553Z graph_break []
2025-12-04T09:45:16.4187626Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4187674Z Autotune Choices Stats:
2025-12-04T09:45:16.4188425Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.4188563Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4188681Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4188840Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4189458Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4190066Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4190713Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4191332Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4191955Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4192573Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4193181Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4193799Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4194407Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4195012Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4195141Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.4195192Z Autotune Choices Stats:
2025-12-04T09:45:16.4195966Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.4196185Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4196363Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4196645Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4197284Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4197924Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4198551Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4199194Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4199841Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4200520Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4201158Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4201793Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4202432Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4203059Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4203187Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.4203261Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4203304Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4203342Z unimplemented []
2025-12-04T09:45:16.4203404Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4203504Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4204079Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4204128Z graph_break []
2025-12-04T09:45:16.4204211Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4204251Z Autotune Choices Stats:
2025-12-04T09:45:16.4205004Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.4205134Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4205250Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4205413Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4208264Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4208875Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4209482Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4210088Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4210771Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4211390Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4211995Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4212616Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4213225Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4213831Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4213963Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.4214005Z Autotune Choices Stats:
2025-12-04T09:45:16.4214771Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.4215005Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4215184Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4215465Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4216108Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4216737Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4217375Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4218018Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4218649Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4219289Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4219940Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4220602Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4221228Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4221880Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4222011Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.4222088Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4222131Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4222169Z unimplemented []
2025-12-04T09:45:16.4222231Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4222333Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4222909Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4222961Z graph_break []
2025-12-04T09:45:16.4223033Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4223075Z Autotune Choices Stats:
2025-12-04T09:45:16.4223846Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.4223977Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4224106Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4224266Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4224876Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4225494Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4226104Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4226712Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4227322Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4227947Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4228563Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4229171Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4229788Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4230394Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4230563Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.4230605Z Autotune Choices Stats:
2025-12-04T09:45:16.4231371Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.4231593Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4231777Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4232054Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4232712Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4233341Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4233968Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4234603Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4235251Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4235901Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4236552Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4237192Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4237821Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4238457Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4238590Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.4238664Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4238708Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4238746Z unimplemented []
2025-12-04T09:45:16.4238807Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4238907Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4239483Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4239519Z graph_break []
2025-12-04T09:45:16.4239595Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4239634Z Autotune Choices Stats:
2025-12-04T09:45:16.4240379Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.4240549Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4240680Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4240842Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4241477Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4242084Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4242702Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4243312Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4243910Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4244525Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4245158Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4245776Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4246382Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4246997Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4247129Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.4247170Z Autotune Choices Stats:
2025-12-04T09:45:16.4247933Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.4248154Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4248321Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4248607Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4249256Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4249885Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4250557Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4251188Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4251819Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4252448Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4253076Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4253730Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4254370Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4255000Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4255139Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.4255215Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4255257Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4255296Z unimplemented []
2025-12-04T09:45:16.4255356Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4255456Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4256041Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4256079Z graph_break []
2025-12-04T09:45:16.4256154Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4256196Z Autotune Choices Stats:
2025-12-04T09:45:16.4256941Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.4257072Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4257198Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4257359Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4257989Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4258606Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4259212Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4259828Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4260474Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4261083Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4261806Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4262447Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4263069Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4263676Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4263818Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.4263861Z Autotune Choices Stats:
2025-12-04T09:45:16.4264633Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.4264855Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4265025Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4265305Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4265944Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4266593Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4267225Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4267850Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4268493Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4269126Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4269753Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4270385Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4271148Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4271783Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4271912Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.4271988Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4272030Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4272069Z unimplemented []
2025-12-04T09:45:16.4272130Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4272243Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4272819Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4272856Z graph_break []
2025-12-04T09:45:16.4272932Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4272972Z Autotune Choices Stats:
2025-12-04T09:45:16.4273739Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.4273869Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4273984Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4274147Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4274770Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4275409Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4276019Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4276623Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4277239Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4277850Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4278463Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4279072Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4279706Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4280322Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4280497Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.4280539Z Autotune Choices Stats:
2025-12-04T09:45:16.4281307Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.4281540Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4281707Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4281990Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4282629Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4283250Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4283893Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4284534Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4285169Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4285809Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4286436Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4287067Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4287697Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4288350Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4288484Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.4288558Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4288611Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4288648Z unimplemented []
2025-12-04T09:45:16.4288709Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4288808Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4289381Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4289443Z graph_break []
2025-12-04T09:45:16.4289518Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4289559Z Autotune Choices Stats:
2025-12-04T09:45:16.4290303Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.4290469Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4290587Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4290748Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4291369Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4291973Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4292605Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4293219Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4293823Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4294443Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4295057Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4295667Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4296285Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4296913Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4297043Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.4297085Z Autotune Choices Stats:
2025-12-04T09:45:16.4297858Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.4298087Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4298254Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4298533Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4299169Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4299799Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4300457Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4301114Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4301759Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4302389Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4303026Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4303654Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4304290Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4304921Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4305061Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.4305144Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4305187Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4305226Z unimplemented []
2025-12-04T09:45:16.4305287Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4305390Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4305985Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4306022Z graph_break []
2025-12-04T09:45:16.4306096Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4306137Z Autotune Choices Stats:
2025-12-04T09:45:16.4306883Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.4307023Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4307138Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4307299Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4307917Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4308530Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4309142Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4309770Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4310386Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4311033Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4311658Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4312268Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4312887Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4313491Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4313635Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.4313675Z Autotune Choices Stats:
2025-12-04T09:45:16.4314464Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.4314689Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4314861Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4315150Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4315792Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4316423Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4317078Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4317737Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4318388Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4319030Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4319657Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4320303Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4320975Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4321604Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4321733Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.4321825Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4321869Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4321906Z unimplemented []
2025-12-04T09:45:16.4321967Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4322066Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4322656Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4322695Z graph_break []
2025-12-04T09:45:16.4322771Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4322812Z Autotune Choices Stats:
2025-12-04T09:45:16.4323566Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.4323709Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4323825Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4323988Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4324598Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4325204Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4325818Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4326436Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4327056Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4327682Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4328292Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4328911Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4329518Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4330143Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4330272Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.4330327Z Autotune Choices Stats:
2025-12-04T09:45:16.4331124Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.4331344Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4331525Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4331800Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4332441Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4333083Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4333707Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4334343Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4334994Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4335641Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4336283Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4336915Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4337554Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4338183Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4338313Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.4338389Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4338432Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4338471Z unimplemented []
2025-12-04T09:45:16.4338532Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4338634Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4339221Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4339277Z graph_break []
2025-12-04T09:45:16.4339352Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4339401Z Autotune Choices Stats:
2025-12-04T09:45:16.4340160Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.4340289Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4340436Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4340598Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4341240Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4341849Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4342458Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4343062Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4343688Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4344304Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4344930Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4345536Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4346155Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4346762Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4346894Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.4346934Z Autotune Choices Stats:
2025-12-04T09:45:16.4347698Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.4347927Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4348102Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4348381Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4349033Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4349654Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4350292Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4350960Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4351588Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4352232Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4352881Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4353531Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4354160Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4354806Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4354935Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.4355011Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4355056Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4355093Z unimplemented []
2025-12-04T09:45:16.4355154Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4355255Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4355831Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4355869Z graph_break []
2025-12-04T09:45:16.4355941Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4355994Z Autotune Choices Stats:
2025-12-04T09:45:16.4356750Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.4356879Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4356993Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4357166Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4357785Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4358403Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4359009Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4359615Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4360224Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4360898Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4361519Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4362128Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4362745Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4363354Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4363485Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.4363527Z Autotune Choices Stats:
2025-12-04T09:45:16.4364284Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.4364504Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4364682Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4364962Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4365606Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4366244Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4366864Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4367498Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4368128Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4368761Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4369407Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4370046Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4370705Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4371345Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4371473Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.4371549Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4371591Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4371630Z unimplemented []
2025-12-04T09:45:16.4371691Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4371792Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4372378Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4372416Z graph_break []
2025-12-04T09:45:16.4372490Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4372530Z Autotune Choices Stats:
2025-12-04T09:45:16.4373280Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.4373420Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4373549Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4373710Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4374340Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4374946Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4375564Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4376168Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4376773Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4377377Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4378004Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4378615Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4379226Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4379842Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4379972Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.4380013Z Autotune Choices Stats:
2025-12-04T09:45:16.4380809Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.4381028Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4381195Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4381476Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4382141Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4382779Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4383405Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4384044Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4384682Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4385317Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4385947Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4386603Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4387238Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4387865Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4388002Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.4388076Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4388118Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4388157Z unimplemented []
2025-12-04T09:45:16.4388218Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4388318Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4388902Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4388940Z graph_break []
2025-12-04T09:45:16.4389013Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4389054Z Autotune Choices Stats:
2025-12-04T09:45:16.4389805Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.4389934Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4390048Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4390221Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4390902Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4391520Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4392123Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4392743Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4393353Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4393961Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4394565Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4395202Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4395817Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4396421Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4396560Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.4396601Z Autotune Choices Stats:
2025-12-04T09:45:16.4397378Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.4397602Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4397769Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4398046Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4398676Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4399325Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4399962Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4400605Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4401250Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4401883Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4402513Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4403145Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4403810Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4404440Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4404568Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.4404642Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4404684Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4404723Z unimplemented []
2025-12-04T09:45:16.4404783Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4404895Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4405467Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4405506Z graph_break []
2025-12-04T09:45:16.4405579Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4405621Z Autotune Choices Stats:
2025-12-04T09:45:16.4406360Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.4406489Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4406604Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4406764Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4407383Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4408008Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4408626Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4409236Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4409850Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4410498Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4411118Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4411735Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4412368Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4412990Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4413120Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.4413161Z Autotune Choices Stats:
2025-12-04T09:45:16.4413933Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.4414163Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4414331Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4414608Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4415252Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4415884Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4416533Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4417173Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4417807Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4418460Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4419102Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4419734Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4420366Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4421058Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4421187Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.4421260Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4421314Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4421352Z unimplemented []
2025-12-04T09:45:16.4421413Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4421512Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4422092Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4422141Z graph_break []
2025-12-04T09:45:16.4422216Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4422255Z Autotune Choices Stats:
2025-12-04T09:45:16.4422997Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.4423126Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4423239Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4423403Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4424025Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4424635Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4425262Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4425877Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4426485Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4427104Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4427707Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4428318Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4428924Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4429550Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4429680Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.4429721Z Autotune Choices Stats:
2025-12-04T09:45:16.4430532Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.4430766Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4430935Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4431215Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4431865Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4432497Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4433131Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4433786Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4434432Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4435061Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4435698Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4436330Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4436961Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4437588Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4437727Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.4437813Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4437856Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4437895Z unimplemented []
2025-12-04T09:45:16.4437956Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4438058Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4438642Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4438681Z graph_break []
2025-12-04T09:45:16.4438754Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4438795Z Autotune Choices Stats:
2025-12-04T09:45:16.4439548Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.4439684Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4439800Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4439962Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4440609Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4441224Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4441830Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4442463Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4443079Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4443690Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4444313Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4444921Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4445537Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4446145Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4446285Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.4446335Z Autotune Choices Stats:
2025-12-04T09:45:16.4447113Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.4447332Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4447499Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4447788Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4448426Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4449056Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4449675Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4450308Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4450989Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4451635Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4452263Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4452905Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4453538Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4454165Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4454293Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.4454378Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4454422Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4454459Z unimplemented []
2025-12-04T09:45:16.4454520Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4454619Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4455210Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4455248Z graph_break []
2025-12-04T09:45:16.4455322Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4455371Z Autotune Choices Stats:
2025-12-04T09:45:16.4456117Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.4456256Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4456373Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4456533Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4457148Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4457760Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4458368Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4458991Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4459609Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4460228Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4460871Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4461493Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4462102Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4462708Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4462839Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.4462892Z Autotune Choices Stats:
2025-12-04T09:45:16.4463669Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.4463891Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4464072Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4464349Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4464994Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4465632Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4466261Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4466892Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4467526Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4468180Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4468819Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4469452Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4470085Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4470753Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4470881Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.4470956Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4470998Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4471036Z unimplemented []
2025-12-04T09:45:16.4471096Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4471196Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4471779Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4471830Z graph_break []
2025-12-04T09:45:16.4471914Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4471955Z Autotune Choices Stats:
2025-12-04T09:45:16.4472712Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.4472842Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4472957Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4473118Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4473747Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4474355Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4474974Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4475578Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4476202Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4476818Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4477431Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4478049Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4478657Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4479269Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4479399Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.4479441Z Autotune Choices Stats:
2025-12-04T09:45:16.4480209Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.4480483Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4480664Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4480941Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4481591Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4482218Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4482857Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4483483Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4484124Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4484777Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4485417Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4486050Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4486690Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4487322Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4487452Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.4487528Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4487570Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4487607Z unimplemented []
2025-12-04T09:45:16.4487668Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4487768Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4488348Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4488396Z graph_break []
2025-12-04T09:45:16.4488469Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4488509Z Autotune Choices Stats:
2025-12-04T09:45:16.4489270Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.4489401Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4489527Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4489687Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4490309Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4490944Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4491556Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4492168Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4492776Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4493408Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4494028Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4494640Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4495259Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4495866Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4495997Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.4496037Z Autotune Choices Stats:
2025-12-04T09:45:16.4496801Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.4497021Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4497196Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4497487Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4498138Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4498771Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4499415Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4500043Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4500712Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4501344Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4501996Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4502633Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4503264Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4503906Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4504036Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.4504111Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4504153Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4504192Z unimplemented []
2025-12-04T09:45:16.4504253Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4504355Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4504929Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4504967Z graph_break []
2025-12-04T09:45:16.4505040Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4505080Z Autotune Choices Stats:
2025-12-04T09:45:16.4505827Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.4505966Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4506090Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4506251Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4506883Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4507494Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4508117Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4508721Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4509331Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4509939Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4510613Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4511238Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4511852Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4512472Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4512602Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.4512643Z Autotune Choices Stats:
2025-12-04T09:45:16.4513407Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.4513627Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4513796Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4514081Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4514751Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4515388Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4516015Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4516653Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4517280Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4517912Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4518548Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4519194Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4519828Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4520478Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4520621Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.4520698Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4520740Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4520778Z unimplemented []
2025-12-04T09:45:16.4520837Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4520938Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4521520Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4521557Z graph_break []
2025-12-04T09:45:16.4521632Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4521671Z Autotune Choices Stats:
2025-12-04T09:45:16.4522421Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:16.4522551Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4522678Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4522841Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4523468Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4524090Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4524702Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4525321Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4525940Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4526541Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4527151Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4527778Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4528399Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4529007Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4529148Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:16.4529188Z Autotune Choices Stats:
2025-12-04T09:45:16.4529949Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:16.4530169Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4530336Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4530653Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4531289Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4531953Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4532592Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4533220Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4533866Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4534499Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4535139Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4535769Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4536420Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4537061Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4537192Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:16.4537264Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4537306Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4537356Z unimplemented []
2025-12-04T09:45:16.4537418Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4537519Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4538102Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4538139Z graph_break []
2025-12-04T09:45:16.4538212Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4538252Z Autotune Choices Stats:
2025-12-04T09:45:16.4538996Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.4539125Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4539241Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4539402Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4540025Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4540691Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4541310Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4541917Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4542539Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4543162Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4543768Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4544376Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4545005Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4545621Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4545751Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:16.4545792Z Autotune Choices Stats:
2025-12-04T09:45:16.4546562Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.4546800Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4546966Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4547247Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4547883Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4548507Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4549158Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4549794Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4550459Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4551110Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4551737Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4552368Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4553013Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4553660Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4553790Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:16.4553877Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4553919Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4553958Z unimplemented []
2025-12-04T09:45:16.4554017Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4554118Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4554697Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4554744Z graph_break []
2025-12-04T09:45:16.4554820Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4554859Z Autotune Choices Stats:
2025-12-04T09:45:16.4555608Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.4555736Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4555852Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4556013Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4556632Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4557254Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4557874Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4558491Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4559095Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4559715Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4560327Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4560972Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4561574Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4562211Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4562363Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:16.4562403Z Autotune Choices Stats:
2025-12-04T09:45:16.4563165Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.4563400Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4563567Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4563851Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4564483Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4565117Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4565745Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4566395Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4567029Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4567658Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4568302Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4568932Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4569565Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4570194Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4570342Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:16.4570445Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4570488Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4570527Z unimplemented []
2025-12-04T09:45:16.4570588Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4570689Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4571291Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4571333Z graph_break []
2025-12-04T09:45:16.4571406Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4571448Z Autotune Choices Stats:
2025-12-04T09:45:16.4572207Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:16.4572336Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4572451Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4572612Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4573232Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4573854Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4574486Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4575098Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4575705Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4576312Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4576936Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4577563Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4578180Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4578799Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4578944Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:16.4578986Z Autotune Choices Stats:
2025-12-04T09:45:16.4579763Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.4579984Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4580150Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4580475Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4581111Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4581743Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4582372Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4583001Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4583655Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4584302Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4584928Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4585570Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4586207Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4586833Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4586970Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:16.4587045Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4587086Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4587124Z unimplemented []
2025-12-04T09:45:16.4587184Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4587285Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4587873Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4587912Z graph_break []
2025-12-04T09:45:16.4587996Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4588036Z Autotune Choices Stats:
2025-12-04T09:45:16.4588785Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:16.4588922Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4589040Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4589204Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4589825Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4590477Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4591101Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4591734Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4592356Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4592964Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4593578Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4594189Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4594803Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4595411Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4595557Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:16.4595597Z Autotune Choices Stats:
2025-12-04T09:45:16.4596373Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.4596603Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4596774Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4597057Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4597692Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4598324Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4598959Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4599580Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4600235Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4600909Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4601533Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4602176Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4602807Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4603453Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4603585Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:16.4603677Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.4603726Z Traceback (most recent call last):
2025-12-04T09:45:16.4603882Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.4603935Z     self.assertTrue(
2025-12-04T09:45:16.4604041Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.4604093Z     raise self.failureException(msg)
2025-12-04T09:45:16.4604218Z AssertionError: False is not true : Log file /tmp/tmp4_r1jh6s/flex_attention_configs.json was not created
2025-12-04T09:45:16.4604221Z 
2025-12-04T09:45:16.4604299Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.4604477Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.4604480Z 
2025-12-04T09:45:16.4604570Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.4604648Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4604692Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4604729Z unimplemented []
2025-12-04T09:45:16.4604791Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4605390Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.4605490Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4605529Z graph_break []
2025-12-04T09:45:16.4605602Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4606111Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.4606161Z   current_size = base.storage().size()
2025-12-04T09:45:16.4606201Z Autotune Choices Stats:
2025-12-04T09:45:16.4606960Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.4607091Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4607207Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4607368Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4607992Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4608617Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4609233Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4609839Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4610485Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4611106Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4611715Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4612323Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4612931Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4613560Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4613703Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.4613744Z Autotune Choices Stats:
2025-12-04T09:45:16.4614511Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.4614741Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4614909Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4615188Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4615821Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4616470Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4617096Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4617738Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4618378Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4619003Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4619636Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4620266Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4620933Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4621566Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4621724Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.4621798Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4621842Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4621880Z unimplemented []
2025-12-04T09:45:16.4621941Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4622040Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4622632Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4622669Z graph_break []
2025-12-04T09:45:16.4622743Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4622795Z Autotune Choices Stats:
2025-12-04T09:45:16.4623544Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.4623674Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4623787Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4623951Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4624566Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4625165Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4625783Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4626401Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4627007Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4627619Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4628224Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4628832Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4629441Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4630053Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4630193Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.4630233Z Autotune Choices Stats:
2025-12-04T09:45:16.4631032Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.4631254Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4631419Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4631716Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4632350Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4632987Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4633622Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4634267Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4634922Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4635558Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4636180Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4636825Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4637454Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4638080Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4638221Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.4638296Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4638338Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4638376Z unimplemented []
2025-12-04T09:45:16.4638435Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4638537Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4639206Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4639245Z graph_break []
2025-12-04T09:45:16.4639328Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4639370Z Autotune Choices Stats:
2025-12-04T09:45:16.4640115Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.4640253Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4640370Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4640568Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4641190Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4641794Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4642404Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4643040Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4643652Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4644263Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4644883Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4645493Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4646102Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4646708Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4646850Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.4646890Z Autotune Choices Stats:
2025-12-04T09:45:16.4647671Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.4647891Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4648068Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4648347Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4648986Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4649614Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4650241Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4650897Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4651547Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4652200Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4652829Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4653467Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4654100Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4654727Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4654858Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.4654932Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4654974Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4655012Z unimplemented []
2025-12-04T09:45:16.4655073Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4655174Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4655761Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4655807Z graph_break []
2025-12-04T09:45:16.4655881Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4655921Z Autotune Choices Stats:
2025-12-04T09:45:16.4656676Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.4656806Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4656922Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4657096Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4657712Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4658323Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4658949Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4659558Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4660188Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4660838Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4661457Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4662854Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4664110Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4665359Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4666139Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.4666345Z Autotune Choices Stats:
2025-12-04T09:45:16.4667173Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.4668223Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4668640Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4669136Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4670087Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4671426Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4672724Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4674016Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4675312Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4676662Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4677969Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4679274Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4680724Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4682017Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4682814Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.4683055Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4683212Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4683323Z unimplemented []
2025-12-04T09:45:16.4683444Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4683643Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4684358Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4687238Z graph_break []
2025-12-04T09:45:16.4687378Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4687534Z Autotune Choices Stats:
2025-12-04T09:45:16.4688371Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.4689274Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4689573Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4689885Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4690734Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4691998Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4693249Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4694498Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4695748Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4697016Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4698275Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4699525Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4700813Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4702059Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4702836Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.4703043Z Autotune Choices Stats:
2025-12-04T09:45:16.4703866Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.4704889Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4705329Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4705820Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4706778Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4708065Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4709359Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4710689Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4711970Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4713256Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4714571Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4715867Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4717175Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4718478Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4719264Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.4719506Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4719660Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4719769Z unimplemented []
2025-12-04T09:45:16.4719889Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4720084Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4720840Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4721483Z graph_break []
2025-12-04T09:45:16.4721612Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4721765Z Autotune Choices Stats:
2025-12-04T09:45:16.4722579Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.4723491Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4723780Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4724095Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4724940Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4726198Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4727452Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4728695Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4729932Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4731212Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4732486Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4733742Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4734992Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4736253Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4737022Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.4737226Z Autotune Choices Stats:
2025-12-04T09:45:16.4738048Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.4739056Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4739477Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4739960Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4740966Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4742273Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4743559Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4744874Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4746187Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4747489Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4748782Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4750094Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4751449Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4752745Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4753553Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.4753789Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4753944Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4754053Z unimplemented []
2025-12-04T09:45:16.4754170Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4754367Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4755083Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4755733Z graph_break []
2025-12-04T09:45:16.4755863Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4756015Z Autotune Choices Stats:
2025-12-04T09:45:16.4756818Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.4757725Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4758016Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4758326Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4759159Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4760448Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4761696Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4762954Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4764199Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4765452Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4766719Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4768017Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4769273Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4770557Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4771351Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.4771569Z Autotune Choices Stats:
2025-12-04T09:45:16.4772404Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.4773418Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4773853Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4774333Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4775274Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4776592Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4777890Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4779186Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4780531Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4781832Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4783126Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4784417Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4785756Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4787072Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4787869Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.4788115Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4788268Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4788391Z unimplemented []
2025-12-04T09:45:16.4788508Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4788704Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4789418Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4790081Z graph_break []
2025-12-04T09:45:16.4790223Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4790384Z Autotune Choices Stats:
2025-12-04T09:45:16.4791240Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.4792159Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4792449Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4792772Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4793605Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4794903Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4796161Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4797399Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4798656Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4799905Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4801197Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4802463Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4803764Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4805039Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4805825Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.4806040Z Autotune Choices Stats:
2025-12-04T09:45:16.4806861Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.4807894Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4808326Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4808816Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4809768Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4811120Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4812443Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4813742Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4815034Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4816341Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4817636Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4818926Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4820221Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4821583Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4822372Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.4822623Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4822779Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4822888Z unimplemented []
2025-12-04T09:45:16.4823005Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4823205Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4823924Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4824589Z graph_break []
2025-12-04T09:45:16.4824717Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4824869Z Autotune Choices Stats:
2025-12-04T09:45:16.4825689Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.4826587Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4826866Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4827178Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4828000Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4829269Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4830576Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4831835Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4833081Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4834335Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4835603Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4836872Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4838114Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4839384Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4840159Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.4840373Z Autotune Choices Stats:
2025-12-04T09:45:16.4841258Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.4842284Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4842704Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4843182Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4844125Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4845412Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4846706Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4848027Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4849328Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4850675Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4851978Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4853274Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4854568Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4855857Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4856666Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.4856904Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4857057Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4857166Z unimplemented []
2025-12-04T09:45:16.4857282Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4857476Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4858213Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4858857Z graph_break []
2025-12-04T09:45:16.4858986Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4859138Z Autotune Choices Stats:
2025-12-04T09:45:16.4859949Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.4860898Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4861177Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4861488Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4862306Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4863556Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4864815Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4866089Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4867348Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4868597Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4869860Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4871146Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4872408Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4873661Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4874455Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.4874660Z Autotune Choices Stats:
2025-12-04T09:45:16.4875499Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.4876599Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4877022Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4877514Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4878469Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4879751Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4881079Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4882351Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4883673Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4884975Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4886276Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4887571Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4888870Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4890178Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4891002Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.4891255Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4891410Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4891519Z unimplemented []
2025-12-04T09:45:16.4891636Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4891835Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4892567Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4893214Z graph_break []
2025-12-04T09:45:16.4893355Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4893505Z Autotune Choices Stats:
2025-12-04T09:45:16.4894309Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.4895216Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4895494Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4895803Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4896616Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4897876Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4899123Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4900394Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4901679Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4902923Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4904174Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4905435Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4906692Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4907941Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4908726Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.4908931Z Autotune Choices Stats:
2025-12-04T09:45:16.4909787Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.4910834Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4911266Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4911748Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4912691Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4913994Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4915303Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4916592Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4917889Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4919212Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4920550Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4921842Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4923135Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4924429Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4925221Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.4925457Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4925610Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4925719Z unimplemented []
2025-12-04T09:45:16.4925836Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4926030Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4926758Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4927398Z graph_break []
2025-12-04T09:45:16.4927539Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4927691Z Autotune Choices Stats:
2025-12-04T09:45:16.4928502Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.4929407Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4929685Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4929996Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4930865Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4932111Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4933362Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4934631Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4935901Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4937163Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4938414Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4939673Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4940966Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4942219Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4942988Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.4943193Z Autotune Choices Stats:
2025-12-04T09:45:16.4944015Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.4945046Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4945476Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4945956Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4946920Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4948211Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4949505Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4950826Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4952118Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4953429Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4954731Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4956024Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4957345Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4958633Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4959426Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.4959666Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4959820Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4959929Z unimplemented []
2025-12-04T09:45:16.4960047Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4960241Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4960984Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.4961641Z graph_break []
2025-12-04T09:45:16.4961770Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4961922Z Autotune Choices Stats:
2025-12-04T09:45:16.4962744Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.4963651Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4963940Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4964248Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4965070Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4966323Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4967562Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4968815Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4970068Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4971380Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4972650Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4973909Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4975164Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4976413Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4977180Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.4977383Z Autotune Choices Stats:
2025-12-04T09:45:16.4978206Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.4979209Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4979639Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4980127Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4981127Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4982420Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4983719Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4985002Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4986299Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4987597Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4988896Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4990219Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.4991557Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4992858Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.4993651Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.4993888Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.4994040Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.4994148Z unimplemented []
2025-12-04T09:45:16.4994268Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.4994460Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.4995180Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.4995821Z graph_break []
2025-12-04T09:45:16.4995952Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.4996103Z Autotune Choices Stats:
2025-12-04T09:45:16.4996925Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.4997844Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.4998133Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.4998444Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.4999276Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5000565Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5001836Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5003084Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5004330Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5005578Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5006851Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5008112Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5009380Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5010697Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5011465Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.5011668Z Autotune Choices Stats:
2025-12-04T09:45:16.5012499Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.5013513Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5013937Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5014415Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5015409Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5016722Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5018008Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5019308Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5020650Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5021946Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5023236Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5024574Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5025898Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5027190Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5027997Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.5028238Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5028391Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5028499Z unimplemented []
2025-12-04T09:45:16.5028616Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5028809Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5029523Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5030172Z graph_break []
2025-12-04T09:45:16.5030302Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5030501Z Autotune Choices Stats:
2025-12-04T09:45:16.5031313Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.5032209Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5032505Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5032814Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5033656Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5034918Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5036167Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5037427Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5038672Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5039922Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5041232Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5042509Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5043765Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5045004Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5045788Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.5045992Z Autotune Choices Stats:
2025-12-04T09:45:16.5046827Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.5047838Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5048260Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5048745Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5049693Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5051047Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5052347Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5053633Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5054952Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5056266Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5057557Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5058869Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5060190Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5061547Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5062337Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.5062575Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5062727Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5062835Z unimplemented []
2025-12-04T09:45:16.5062965Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5063158Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5063874Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5064516Z graph_break []
2025-12-04T09:45:16.5064645Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5064796Z Autotune Choices Stats:
2025-12-04T09:45:16.5065605Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.5066509Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5066788Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5067097Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5067911Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5069191Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5070494Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5071754Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5073012Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5074264Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5075512Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5076757Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5078050Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5079300Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5080074Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.5080279Z Autotune Choices Stats:
2025-12-04T09:45:16.5081151Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.5082184Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5082601Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5083087Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5084037Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5085331Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5086648Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5087933Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5089236Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5090599Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5091891Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5093180Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5094478Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5095794Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5096582Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.5096831Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5096874Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5096919Z unimplemented []
2025-12-04T09:45:16.5096980Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5097083Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5097659Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5097711Z graph_break []
2025-12-04T09:45:16.5097787Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5097828Z Autotune Choices Stats:
2025-12-04T09:45:16.5098582Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.5098711Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5098829Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5098992Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5099609Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5100301Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5100967Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5101584Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5102190Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5102812Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5103424Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5104032Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5104639Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5105260Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5105393Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.5105434Z Autotune Choices Stats:
2025-12-04T09:45:16.5106209Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.5106437Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5106607Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5106886Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5107515Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5108137Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5108763Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5109407Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5110049Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5110707Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5111350Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5111983Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5112628Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5113250Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5113393Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.5113479Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5113526Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5113563Z unimplemented []
2025-12-04T09:45:16.5113625Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5113724Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5114307Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5114346Z graph_break []
2025-12-04T09:45:16.5114420Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5114462Z Autotune Choices Stats:
2025-12-04T09:45:16.5115219Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.5115360Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5115477Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5115642Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5116264Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5116872Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5117480Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5118109Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5118727Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5119331Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5119949Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5120596Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5121206Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5121803Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5121949Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.5122002Z Autotune Choices Stats:
2025-12-04T09:45:16.5122774Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.5122994Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5123161Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5123453Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5124091Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5124720Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5125348Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5125970Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5126627Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5127269Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5127886Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5128525Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5129161Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5129792Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5129920Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.5130005Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5130048Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5130087Z unimplemented []
2025-12-04T09:45:16.5130148Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5130250Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5130869Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5130909Z graph_break []
2025-12-04T09:45:16.5130983Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5131034Z Autotune Choices Stats:
2025-12-04T09:45:16.5131780Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.5131925Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5132043Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5132203Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5132815Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5133427Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5134052Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5134669Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5135278Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5135890Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5136498Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5137120Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5137735Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5138348Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5138480Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.5138534Z Autotune Choices Stats:
2025-12-04T09:45:16.5139314Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.5139533Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5139713Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5139997Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5140656Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5141300Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5141929Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5142559Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5143185Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5143847Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5144486Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5145110Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5145744Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5146371Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5146500Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.5146573Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5146616Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5146653Z unimplemented []
2025-12-04T09:45:16.5146713Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5146813Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5147389Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5147435Z graph_break []
2025-12-04T09:45:16.5147517Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5147560Z Autotune Choices Stats:
2025-12-04T09:45:16.5148314Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.5148446Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5148560Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5148722Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5149352Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5149961Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5150607Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5151222Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5151855Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5152470Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5153087Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5153693Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5154316Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5154927Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5155057Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.5155099Z Autotune Choices Stats:
2025-12-04T09:45:16.5155859Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.5156088Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5156262Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5156545Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5157190Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5157816Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5158458Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5159086Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5159717Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5160356Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5161031Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5161679Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5162308Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5162947Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5163076Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.5163152Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5163194Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5163233Z unimplemented []
2025-12-04T09:45:16.5163295Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5163396Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5163977Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5164016Z graph_break []
2025-12-04T09:45:16.5164103Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5164146Z Autotune Choices Stats:
2025-12-04T09:45:16.5164906Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.5165034Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5165161Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5165320Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5165935Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5166553Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5167165Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5167772Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5168371Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5169003Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5169620Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5170228Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5170873Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5171480Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5171610Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.5171653Z Autotune Choices Stats:
2025-12-04T09:45:16.5172419Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.5172637Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5172823Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5173101Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5173755Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5174398Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5175024Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5175661Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5176295Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5176929Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5177579Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5178219Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5178849Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5179485Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5179615Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.5179690Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5179733Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5179770Z unimplemented []
2025-12-04T09:45:16.5179831Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5179930Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5180545Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5180583Z graph_break []
2025-12-04T09:45:16.5180658Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5180697Z Autotune Choices Stats:
2025-12-04T09:45:16.5181440Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.5181585Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5181712Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5181876Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5182511Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5183118Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5183735Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5184344Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5184962Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5185569Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5186200Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5186816Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5187426Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5188044Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5188177Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.5188217Z Autotune Choices Stats:
2025-12-04T09:45:16.5188988Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.5189210Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5189378Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5189656Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5190321Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5190998Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5191633Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5192274Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5192911Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5193541Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5194168Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5194825Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5195466Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5196093Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5196236Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.5196312Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5196354Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5196392Z unimplemented []
2025-12-04T09:45:16.5196452Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5196552Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5197133Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5197174Z graph_break []
2025-12-04T09:45:16.5197248Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5197289Z Autotune Choices Stats:
2025-12-04T09:45:16.5198042Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.5198170Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5198300Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5198463Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5199095Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5199714Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5200326Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5200995Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5201607Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5202216Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5202831Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5203492Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5204123Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5204729Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5204880Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.5204921Z Autotune Choices Stats:
2025-12-04T09:45:16.5205679Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.5205898Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5206067Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5206345Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5206986Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5207639Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5208276Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5208904Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5209548Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5210178Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5210845Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5211468Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5212132Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5212767Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5212897Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.5212971Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5213014Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5213052Z unimplemented []
2025-12-04T09:45:16.5213127Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5213226Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5213803Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5213840Z graph_break []
2025-12-04T09:45:16.5213915Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5213955Z Autotune Choices Stats:
2025-12-04T09:45:16.5214709Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.5214837Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5214953Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5215115Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5215727Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5216358Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5216977Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5217588Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5218206Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5218817Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5219429Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5220041Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5220698Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5221315Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5221445Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.5221486Z Autotune Choices Stats:
2025-12-04T09:45:16.5222251Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.5222487Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5222655Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5222945Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5223585Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5226598Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5227264Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5227907Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5228543Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5229184Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5229818Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5230478Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5231109Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5231762Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5231895Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.5231987Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5232030Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5232070Z unimplemented []
2025-12-04T09:45:16.5232132Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5232235Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5232822Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5232875Z graph_break []
2025-12-04T09:45:16.5232949Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5232991Z Autotune Choices Stats:
2025-12-04T09:45:16.5233739Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.5233868Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5233988Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5234153Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5234776Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5235391Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5236019Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5236643Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5237251Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5237870Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5238492Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5239111Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5239723Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5240352Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5240523Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.5240566Z Autotune Choices Stats:
2025-12-04T09:45:16.5241349Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.5241582Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5241755Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5242036Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5242675Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5243300Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5243945Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5244594Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5245239Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5245867Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5246507Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5247137Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5247769Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5248400Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5248538Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.5248626Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5248669Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5248708Z unimplemented []
2025-12-04T09:45:16.5248770Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5248873Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5249460Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5249497Z graph_break []
2025-12-04T09:45:16.5249572Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5249612Z Autotune Choices Stats:
2025-12-04T09:45:16.5250365Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.5250548Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5250664Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5250827Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5251437Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5252043Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5252668Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5253285Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5253905Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5254515Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5255139Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5255751Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5256372Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5256975Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5257129Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.5257170Z Autotune Choices Stats:
2025-12-04T09:45:16.5257945Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.5258166Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5258336Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5258624Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5259259Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5259889Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5260552Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5261179Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5261834Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5262480Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5263109Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5263753Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5264384Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5265013Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5265142Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.5265227Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5265271Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5265308Z unimplemented []
2025-12-04T09:45:16.5265371Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5265471Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5266059Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5266099Z graph_break []
2025-12-04T09:45:16.5266182Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5266223Z Autotune Choices Stats:
2025-12-04T09:45:16.5266977Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.5267114Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5267233Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5267396Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5268015Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5268626Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5269229Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5269862Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5270585Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5271194Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5271798Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5272420Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5273027Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5273631Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5273778Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.5273821Z Autotune Choices Stats:
2025-12-04T09:45:16.5274605Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.5274824Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5275001Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5275283Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5275919Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5276559Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5277186Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5277814Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5278458Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5279100Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5279745Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5280376Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5281058Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5281685Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5281815Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.5281889Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5281931Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5281971Z unimplemented []
2025-12-04T09:45:16.5282031Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5282133Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5282736Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5282773Z graph_break []
2025-12-04T09:45:16.5282860Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5282901Z Autotune Choices Stats:
2025-12-04T09:45:16.5283657Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.5283787Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5283902Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5284064Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5284702Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5285312Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5285927Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5286535Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5287171Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5287790Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5288409Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5289901Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5290557Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5291176Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5291309Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.5291350Z Autotune Choices Stats:
2025-12-04T09:45:16.5292123Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.5292375Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5292544Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5292827Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5293461Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5294091Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5294770Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5295392Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5296028Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5296684Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5297315Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5297944Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5298597Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5299229Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5299361Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.5299434Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5299478Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5299516Z unimplemented []
2025-12-04T09:45:16.5299577Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5299679Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5300262Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5300312Z graph_break []
2025-12-04T09:45:16.5300386Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5300461Z Autotune Choices Stats:
2025-12-04T09:45:16.5301241Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.5301371Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5301487Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5301650Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5302271Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5302918Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5303536Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5304146Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5304755Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5305395Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5306005Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5306616Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5307249Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5307856Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5307990Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.5308030Z Autotune Choices Stats:
2025-12-04T09:45:16.5308805Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.5309027Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5309204Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5309493Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5310133Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5310806Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5311462Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5312089Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5312726Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5313355Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5314018Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5314653Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5315289Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5315934Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5316064Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.5316138Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5316180Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5316219Z unimplemented []
2025-12-04T09:45:16.5316280Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5316380Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5316976Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5317015Z graph_break []
2025-12-04T09:45:16.5317087Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5317128Z Autotune Choices Stats:
2025-12-04T09:45:16.5317875Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:16.5318012Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5318138Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5318302Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5318920Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5319532Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5320162Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5320799Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5321414Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5322020Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5322672Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5323282Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5323893Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5324529Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5324659Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:16.5324701Z Autotune Choices Stats:
2025-12-04T09:45:16.5325475Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:16.5325696Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5325862Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5326140Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5326801Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5327432Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5328061Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5328710Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5329346Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5329980Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5330649Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5331316Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5331944Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5332576Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5332729Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:16.5332803Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5332847Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5332884Z unimplemented []
2025-12-04T09:45:16.5332964Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5333066Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5333648Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5333686Z graph_break []
2025-12-04T09:45:16.5333762Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5333801Z Autotune Choices Stats:
2025-12-04T09:45:16.5334549Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.5334676Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5334809Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5334971Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5335601Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5336212Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5336827Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5337460Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5338070Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5338680Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5339315Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5339950Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5340596Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5341201Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5341354Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:16.5341394Z Autotune Choices Stats:
2025-12-04T09:45:16.5342165Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.5342386Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5342554Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5342834Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5343465Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5344119Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5344748Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5345377Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5346034Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5346663Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5347297Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5347935Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5348587Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5349218Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5349349Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:16.5349423Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5349475Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5349514Z unimplemented []
2025-12-04T09:45:16.5349574Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5349674Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5350269Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5350308Z graph_break []
2025-12-04T09:45:16.5350381Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5350463Z Autotune Choices Stats:
2025-12-04T09:45:16.5351216Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.5351344Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5351459Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5351622Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5352238Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5352877Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5353489Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5354101Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5354741Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5355350Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5355963Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5356575Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5357207Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5357817Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5357947Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:16.5357988Z Autotune Choices Stats:
2025-12-04T09:45:16.5358757Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.5358989Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5359157Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5359445Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5360085Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5360758Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5361413Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5362040Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5362672Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5363332Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5363964Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5364593Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5365224Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5365876Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5366007Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:16.5366079Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5366121Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5366159Z unimplemented []
2025-12-04T09:45:16.5366221Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5366322Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5366898Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5366946Z graph_break []
2025-12-04T09:45:16.5367019Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5367060Z Autotune Choices Stats:
2025-12-04T09:45:16.5367822Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:16.5367952Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5368065Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5368229Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5368846Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5369465Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5370093Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5370744Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5371348Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5371984Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5372597Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5373210Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5373818Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5374443Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5374573Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:16.5374613Z Autotune Choices Stats:
2025-12-04T09:45:16.5375380Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.5375610Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5375790Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5376074Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5376711Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5377344Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5377980Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5378631Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5379265Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5379897Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5380580Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5381215Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5381849Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5382491Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5382632Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:16.5382706Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5382748Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5382785Z unimplemented []
2025-12-04T09:45:16.5382845Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5382947Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5383522Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5383561Z graph_break []
2025-12-04T09:45:16.5383634Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5383687Z Autotune Choices Stats:
2025-12-04T09:45:16.5384450Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:16.5384581Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5384695Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5384857Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5385484Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5386111Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5386753Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5387360Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5387962Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5388579Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5389202Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5389816Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5390468Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5391099Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5391228Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:16.5391268Z Autotune Choices Stats:
2025-12-04T09:45:16.5392037Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.5392258Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5392425Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5392717Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5393374Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5394004Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5394630Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5395267Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5395904Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5396536Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5397161Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5397805Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5398437Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5399065Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5399202Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:16.5399277Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5399319Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5399357Z unimplemented []
2025-12-04T09:45:16.5399419Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5399529Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5400108Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5400146Z graph_break []
2025-12-04T09:45:16.5400219Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5400259Z Autotune Choices Stats:
2025-12-04T09:45:16.5401045Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:16.5401188Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5401302Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5401484Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5402096Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5402703Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5403316Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5403947Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5404550Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5405164Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5405793Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5406405Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5407018Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5407626Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5407765Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:16.5407805Z Autotune Choices Stats:
2025-12-04T09:45:16.5408585Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:16.5408805Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5408971Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5409253Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5409896Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5410573Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5411200Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5411829Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5412485Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5413112Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5413740Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5414392Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5415024Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5415656Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5415785Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:16.5415876Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.5415926Z Traceback (most recent call last):
2025-12-04T09:45:16.5416084Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.5416134Z     self.assertTrue(
2025-12-04T09:45:16.5416240Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.5416288Z     raise self.failureException(msg)
2025-12-04T09:45:16.5416417Z AssertionError: False is not true : Log file /tmp/tmpsimzr412/flex_attention_configs.json was not created
2025-12-04T09:45:16.5416420Z 
2025-12-04T09:45:16.5416496Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.5416674Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.5416677Z 
2025-12-04T09:45:16.5416765Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.5416842Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5416884Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5416922Z unimplemented []
2025-12-04T09:45:16.5416985Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5417564Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.5417663Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5417700Z graph_break []
2025-12-04T09:45:16.5417791Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5418289Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.5418339Z   current_size = base.storage().size()
2025-12-04T09:45:16.5418389Z Autotune Choices Stats:
2025-12-04T09:45:16.5419140Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.5419270Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5419385Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5419550Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5420167Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5420847Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5421450Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5422053Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5422671Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5423288Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5423896Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5424501Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5425124Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5425737Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5425871Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.5425911Z Autotune Choices Stats:
2025-12-04T09:45:16.5426679Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.5426912Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5427089Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5427372Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5428003Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5428647Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5429276Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5429916Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5430572Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5431199Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5431875Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5432502Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5433134Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5433813Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5433942Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.5434016Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5434059Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5434096Z unimplemented []
2025-12-04T09:45:16.5434157Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5434257Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5434833Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5434871Z graph_break []
2025-12-04T09:45:16.5434944Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5434998Z Autotune Choices Stats:
2025-12-04T09:45:16.5435755Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.5435886Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5436001Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5436164Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5436779Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5437375Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5438000Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5438606Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5439216Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5439829Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5440502Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5441107Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5441712Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5442345Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5442475Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.5442516Z Autotune Choices Stats:
2025-12-04T09:45:16.5443280Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.5443501Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5443666Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5443969Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5444617Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5445247Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5445871Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5446505Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5447142Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5447772Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5448391Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5449037Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5449667Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5450288Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5450468Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.5450543Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5450585Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5450626Z unimplemented []
2025-12-04T09:45:16.5450686Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5450812Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5451393Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5451432Z graph_break []
2025-12-04T09:45:16.5451506Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5451545Z Autotune Choices Stats:
2025-12-04T09:45:16.5452293Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.5452442Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5452557Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5452729Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5453349Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5453958Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5454564Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5455190Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5455795Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5456406Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5457030Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5457638Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5458250Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5458859Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5458998Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.5459038Z Autotune Choices Stats:
2025-12-04T09:45:16.5459811Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.5460030Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5460198Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5460510Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5461145Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5461797Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5462424Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5463050Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5463707Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5464333Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5464972Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5465624Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5466252Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5466882Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5467013Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.5467086Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5467129Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5467165Z unimplemented []
2025-12-04T09:45:16.5467226Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5467337Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5467921Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5467959Z graph_break []
2025-12-04T09:45:16.5468031Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5468072Z Autotune Choices Stats:
2025-12-04T09:45:16.5468817Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.5468947Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5469061Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5469239Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5469869Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5470509Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5471122Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5471740Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5472375Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5472979Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5473590Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5474226Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5474834Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5475443Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5475572Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.5475613Z Autotune Choices Stats:
2025-12-04T09:45:16.5476367Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.5476608Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5476774Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5477058Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5477702Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5478352Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5478981Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5479605Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5480237Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5480929Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5481554Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5482185Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5482834Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5483463Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5483592Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.5483667Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5483709Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5483748Z unimplemented []
2025-12-04T09:45:16.5483808Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5483909Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5484486Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5484539Z graph_break []
2025-12-04T09:45:16.5484613Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5484653Z Autotune Choices Stats:
2025-12-04T09:45:16.5485415Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.5485542Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5485658Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5485821Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5486438Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5487064Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5487666Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5488273Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5488871Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5489501Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5490113Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5490758Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5491388Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5491995Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5492125Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.5492165Z Autotune Choices Stats:
2025-12-04T09:45:16.5492919Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.5493137Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5493319Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5493616Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5494251Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5494883Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5495528Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5496154Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5496787Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5497416Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5498066Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5498688Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5499316Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5499961Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5500092Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.5500165Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5500208Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5500245Z unimplemented []
2025-12-04T09:45:16.5500306Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5500443Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5501014Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5501051Z graph_break []
2025-12-04T09:45:16.5501124Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5501165Z Autotune Choices Stats:
2025-12-04T09:45:16.5501908Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.5502070Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5502183Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5502347Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5502968Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5503572Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5504202Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5504810Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5505418Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5506023Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5506650Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5507252Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5507866Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5508490Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5508620Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.5508660Z Autotune Choices Stats:
2025-12-04T09:45:16.5509427Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.5509648Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5509814Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5510092Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5510799Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5511426Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5512055Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5512704Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5513329Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5513962Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5514589Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5515250Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5515880Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5516510Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5516650Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.5516725Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5516767Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5516806Z unimplemented []
2025-12-04T09:45:16.5516877Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5516978Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5517556Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5517594Z graph_break []
2025-12-04T09:45:16.5517667Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5517707Z Autotune Choices Stats:
2025-12-04T09:45:16.5518457Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.5518596Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5518712Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5518874Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5519500Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5520109Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5520748Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5521378Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5521984Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5522590Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5523198Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5523824Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5524431Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5525040Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5525179Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.5525220Z Autotune Choices Stats:
2025-12-04T09:45:16.5525996Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.5526215Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5526383Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5526665Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5527302Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5527953Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5528578Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5529202Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5529856Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5530522Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5531152Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5531798Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5532448Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5533073Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5533205Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.5533277Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5533333Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5533371Z unimplemented []
2025-12-04T09:45:16.5533432Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5533533Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5534127Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5534165Z graph_break []
2025-12-04T09:45:16.5534239Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5534279Z Autotune Choices Stats:
2025-12-04T09:45:16.5535029Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.5535158Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5535272Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5535436Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5536053Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5536678Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5537289Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5537898Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5538525Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5539130Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5539742Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5540356Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5541010Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5541616Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5541747Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.5541786Z Autotune Choices Stats:
2025-12-04T09:45:16.5542566Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.5542801Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5542967Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5543250Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5543884Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5544512Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5545158Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5545781Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5546412Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5547068Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5547691Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5548326Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5548959Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5549605Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5549736Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.5549811Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5549852Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5549890Z unimplemented []
2025-12-04T09:45:16.5549952Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5550052Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5550648Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5550706Z graph_break []
2025-12-04T09:45:16.5550779Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5550819Z Autotune Choices Stats:
2025-12-04T09:45:16.5551577Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.5551708Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5551823Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5551982Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5552591Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5553212Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5553838Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5554448Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5555056Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5555683Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5556291Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5556908Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5557519Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5558149Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5558278Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.5558319Z Autotune Choices Stats:
2025-12-04T09:45:16.5559080Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.5559308Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5559486Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5559767Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5560460Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5561092Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5561718Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5562386Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5563019Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5563652Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5564303Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5564937Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5565567Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5566194Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5566345Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.5566419Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5566463Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5566500Z unimplemented []
2025-12-04T09:45:16.5566562Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5566663Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5567242Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5567281Z graph_break []
2025-12-04T09:45:16.5567355Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5567395Z Autotune Choices Stats:
2025-12-04T09:45:16.5568158Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.5568288Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5568402Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5568562Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5569183Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5569782Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5570402Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5571036Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5571646Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5572243Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5572884Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5573494Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5574107Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5574728Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5574873Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.5574913Z Autotune Choices Stats:
2025-12-04T09:45:16.5575679Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.5575902Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5576069Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5576360Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5577010Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5577637Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5578270Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5578898Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5579557Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5580188Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5580852Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5581522Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5582152Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5582780Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5582931Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.5583004Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5583047Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5583084Z unimplemented []
2025-12-04T09:45:16.5583146Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5583246Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5583834Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5583874Z graph_break []
2025-12-04T09:45:16.5583946Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5583987Z Autotune Choices Stats:
2025-12-04T09:45:16.5584737Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.5584876Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5584991Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5585161Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5585781Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5586390Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5587002Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5587627Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5588233Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5588842Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5589454Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5590072Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5590720Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5591326Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5591480Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.5591521Z Autotune Choices Stats:
2025-12-04T09:45:16.5592300Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.5592519Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5592687Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5592968Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5593608Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5594268Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5594894Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5595523Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5596175Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5596802Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5597428Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5598068Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5598708Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5599339Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5599468Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.5599541Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5599582Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5599621Z unimplemented []
2025-12-04T09:45:16.5599681Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5599781Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5600371Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5600458Z graph_break []
2025-12-04T09:45:16.5600532Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5600572Z Autotune Choices Stats:
2025-12-04T09:45:16.5601311Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.5601439Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5601555Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5601729Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5602357Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5602961Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5603564Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5604172Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5604805Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5605408Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5606015Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5606647Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5607255Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5607857Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5607988Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.5608028Z Autotune Choices Stats:
2025-12-04T09:45:16.5608793Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.5609039Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5609204Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5609494Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5610124Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5610788Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5611439Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5612065Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5612699Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5613358Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5613984Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5614617Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5615271Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5615897Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5616029Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.5616104Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5616146Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5616185Z unimplemented []
2025-12-04T09:45:16.5616247Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5616346Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5616919Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5616969Z graph_break []
2025-12-04T09:45:16.5617041Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5617082Z Autotune Choices Stats:
2025-12-04T09:45:16.5617839Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.5617969Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5618085Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5618245Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5618862Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5619488Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5620094Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5620747Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5621352Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5621989Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5622600Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5623210Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5623842Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5624451Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5624582Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.5624623Z Autotune Choices Stats:
2025-12-04T09:45:16.5625389Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.5625610Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5625788Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5626082Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5626719Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5627350Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5627991Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5628617Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5629250Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5629885Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5630573Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5631201Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5631833Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5632493Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5632623Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.5632699Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5632740Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5632779Z unimplemented []
2025-12-04T09:45:16.5632839Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5632940Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5633515Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5633552Z graph_break []
2025-12-04T09:45:16.5633626Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5633665Z Autotune Choices Stats:
2025-12-04T09:45:16.5634416Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.5634561Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5634688Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5634850Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5635464Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5636071Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5636700Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5637301Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5637907Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5638510Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5639141Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5639746Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5640354Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5641022Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5641154Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.5641194Z Autotune Choices Stats:
2025-12-04T09:45:16.5641963Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.5642184Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5642351Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5642636Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5643299Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5643929Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5644556Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5645203Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5645834Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5646468Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5647092Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5647745Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5648368Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5648987Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5649129Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.5649202Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5649245Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5649282Z unimplemented []
2025-12-04T09:45:16.5649355Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5649455Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5650027Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5650067Z graph_break []
2025-12-04T09:45:16.5650139Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5650179Z Autotune Choices Stats:
2025-12-04T09:45:16.5650968Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.5651096Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5651224Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5651388Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5652015Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5652627Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5653237Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5653867Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5654465Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5655074Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5655676Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5656304Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5656915Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5657524Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5657663Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.5657704Z Autotune Choices Stats:
2025-12-04T09:45:16.5658486Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.5658706Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5658878Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5659159Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5659788Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5660475Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5661100Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5661720Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5662376Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5663005Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5663634Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5664266Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5664917Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5665549Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5665679Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.5665754Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5665797Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5665845Z unimplemented []
2025-12-04T09:45:16.5665905Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5666005Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5666588Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5666626Z graph_break []
2025-12-04T09:45:16.5666700Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5666739Z Autotune Choices Stats:
2025-12-04T09:45:16.5667483Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.5667611Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5667729Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5667890Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5668506Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5669129Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5669748Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5670351Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5671020Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5671617Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5672221Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5672822Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5673446Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5674056Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5674186Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.5674227Z Autotune Choices Stats:
2025-12-04T09:45:16.5675006Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.5675237Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5675403Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5675684Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5676327Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5676975Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5677620Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5678249Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5678879Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5679528Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5680151Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5680813Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5681550Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5682209Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5682343Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.5682417Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5682461Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5682498Z unimplemented []
2025-12-04T09:45:16.5682563Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5682662Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5683242Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5683293Z graph_break []
2025-12-04T09:45:16.5683368Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5683410Z Autotune Choices Stats:
2025-12-04T09:45:16.5684175Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.5684305Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5684421Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5684585Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5685206Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5685810Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5686438Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5687045Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5687651Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5688276Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5688889Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5689505Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5690137Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5690810Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5690942Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.5690983Z Autotune Choices Stats:
2025-12-04T09:45:16.5691751Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.5691987Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5692168Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5692450Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5693086Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5693735Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5694366Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5695015Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5695647Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5696276Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5696925Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5697556Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5698192Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5698821Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5698968Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.5699043Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5699085Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5699124Z unimplemented []
2025-12-04T09:45:16.5699184Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5699285Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5699862Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5699901Z graph_break []
2025-12-04T09:45:16.5699975Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5700016Z Autotune Choices Stats:
2025-12-04T09:45:16.5700813Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.5700953Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5701072Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5701233Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5701850Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5702458Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5703079Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5703696Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5704302Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5704908Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5705540Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5706148Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5706760Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5707366Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5707521Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.5707564Z Autotune Choices Stats:
2025-12-04T09:45:16.5708326Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.5708547Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5708715Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5709003Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5709651Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5710275Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5710937Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5711569Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5712231Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5712860Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5713487Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5714141Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5714772Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5715401Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5715531Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.5715614Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5715659Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5715697Z unimplemented []
2025-12-04T09:45:16.5715760Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5715861Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5716454Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5716495Z graph_break []
2025-12-04T09:45:16.5716568Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5716607Z Autotune Choices Stats:
2025-12-04T09:45:16.5717352Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.5717492Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5717609Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5717780Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5718401Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5719008Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5719622Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5720249Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5720881Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5721494Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5722121Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5722763Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5723370Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5723972Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5724115Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.5724155Z Autotune Choices Stats:
2025-12-04T09:45:16.5724934Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.5725156Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5725323Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5725601Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5726240Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5726886Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5727514Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5728145Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5728771Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5729420Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5730051Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5730716Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5731367Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5732001Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5732131Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.5732207Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5732249Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5732288Z unimplemented []
2025-12-04T09:45:16.5732348Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5734659Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5735278Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5735317Z graph_break []
2025-12-04T09:45:16.5735410Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5735453Z Autotune Choices Stats:
2025-12-04T09:45:16.5736210Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.5736344Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5736462Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5736628Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5737270Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5737882Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5738497Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5739114Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5739739Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5740346Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5741008Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5741641Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5742276Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5742880Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5743013Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.5743055Z Autotune Choices Stats:
2025-12-04T09:45:16.5743812Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.5744045Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5744231Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5744509Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5745149Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5745782Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5746429Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5747052Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5747685Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5748334Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5748958Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5749589Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5750233Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5750924Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5751056Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.5751135Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5751179Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5751217Z unimplemented []
2025-12-04T09:45:16.5751280Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5751382Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5751962Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5752018Z graph_break []
2025-12-04T09:45:16.5752092Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5752132Z Autotune Choices Stats:
2025-12-04T09:45:16.5752892Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.5753021Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5753137Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5753301Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5753924Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5754557Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5755169Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5755792Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5756399Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5757037Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5757646Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5758271Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5758895Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5759504Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5759635Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.5759675Z Autotune Choices Stats:
2025-12-04T09:45:16.5760468Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.5760689Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5760871Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5761164Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5761800Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5762431Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5763064Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5763707Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5764336Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5764961Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5765606Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5766233Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5766871Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5767520Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5767650Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.5767725Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5767767Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5767805Z unimplemented []
2025-12-04T09:45:16.5767867Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5767968Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5768549Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5768587Z graph_break []
2025-12-04T09:45:16.5768660Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5768700Z Autotune Choices Stats:
2025-12-04T09:45:16.5769441Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.5769581Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5769707Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5769869Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5770513Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5771118Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5771749Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5772357Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5772963Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5773566Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5774205Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5774808Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5775420Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5776049Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5776181Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.5776222Z Autotune Choices Stats:
2025-12-04T09:45:16.5776984Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.5777205Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5777379Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5777660Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5778320Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5778948Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5779570Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5780217Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5780884Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5781513Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5782161Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5782820Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5783453Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5784079Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5784222Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.5784298Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5784339Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5784377Z unimplemented []
2025-12-04T09:45:16.5784436Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5784552Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5785131Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5785169Z graph_break []
2025-12-04T09:45:16.5785243Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5785282Z Autotune Choices Stats:
2025-12-04T09:45:16.5786040Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.5786168Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5786292Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5786453Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5787081Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5787685Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5788294Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5788931Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5789535Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5790136Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5790777Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5791411Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5792026Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5792645Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5792788Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.5792827Z Autotune Choices Stats:
2025-12-04T09:45:16.5793603Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.5793824Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5793990Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5794270Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5794902Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5795554Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5796178Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5796818Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5797473Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5798102Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5798725Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5799351Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5800004Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5800664Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5800795Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.5800868Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5800911Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5800963Z unimplemented []
2025-12-04T09:45:16.5801023Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5801124Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5801720Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5801758Z graph_break []
2025-12-04T09:45:16.5801832Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5801872Z Autotune Choices Stats:
2025-12-04T09:45:16.5802621Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.5802749Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5802865Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5803028Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5803648Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5804281Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5804884Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5805492Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5806129Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5806737Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5807349Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5807953Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5808580Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5809186Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5809317Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.5809359Z Autotune Choices Stats:
2025-12-04T09:45:16.5810127Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.5810357Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5810554Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5810836Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5811473Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5812104Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5812759Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5813391Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5814023Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5814679Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5815307Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5815945Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5816573Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5817218Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5817349Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.5817425Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5817467Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5817506Z unimplemented []
2025-12-04T09:45:16.5817567Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5817670Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5818247Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5818297Z graph_break []
2025-12-04T09:45:16.5818372Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5818412Z Autotune Choices Stats:
2025-12-04T09:45:16.5819172Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.5819302Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5819417Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5819580Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5820200Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5820853Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5821493Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5822103Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5822709Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5823336Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5823947Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5824561Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5825174Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5825801Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5825934Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.5825974Z Autotune Choices Stats:
2025-12-04T09:45:16.5826740Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.5826969Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5827147Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5827425Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5828058Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5828689Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5829319Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5829963Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5830621Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5831260Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5831913Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5832543Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5833175Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5833806Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5833964Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.5834038Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5834082Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5834120Z unimplemented []
2025-12-04T09:45:16.5834183Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5834283Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5834857Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5834898Z graph_break []
2025-12-04T09:45:16.5834971Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5835012Z Autotune Choices Stats:
2025-12-04T09:45:16.5835788Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.5835918Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5836035Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5836197Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5836817Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5837423Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5838062Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5838667Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5839272Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5839876Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5840549Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5841161Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5841929Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5842550Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5842696Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.5842738Z Autotune Choices Stats:
2025-12-04T09:45:16.5843509Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.5843730Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5843897Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5844189Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5844945Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5845578Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5846203Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5846848Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5847503Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5848138Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5848761Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5849414Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5850049Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5850721Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5850862Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.5850938Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5850980Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5851020Z unimplemented []
2025-12-04T09:45:16.5851082Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5851183Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5851777Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5851819Z graph_break []
2025-12-04T09:45:16.5851893Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5851933Z Autotune Choices Stats:
2025-12-04T09:45:16.5852686Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.5852829Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5852946Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5853129Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5853745Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5854358Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5854967Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5855592Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5856196Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5856819Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5857443Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5858062Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5858669Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5859277Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5859418Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.5859458Z Autotune Choices Stats:
2025-12-04T09:45:16.5860242Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.5860487Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5860657Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5860939Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5861570Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5862222Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5862852Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5863500Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5864155Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5864785Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5865414Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5866046Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5866678Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5867305Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5867437Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.5867511Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5867556Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5867594Z unimplemented []
2025-12-04T09:45:16.5867656Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5867754Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5868344Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5868391Z graph_break []
2025-12-04T09:45:16.5868465Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5868506Z Autotune Choices Stats:
2025-12-04T09:45:16.5869250Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.5869379Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5869494Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5869676Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5870306Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5870947Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5871560Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5872173Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5872915Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5873519Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5874147Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5874787Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5875394Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5876017Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5876146Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.5876187Z Autotune Choices Stats:
2025-12-04T09:45:16.5876955Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.5877202Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5877369Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5877655Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5878310Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5878951Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5879605Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5880227Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5880916Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5881573Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5882201Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5882842Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5883495Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5884120Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5884248Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.5884322Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5884364Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5884402Z unimplemented []
2025-12-04T09:45:16.5884463Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5884564Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5885139Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5885196Z graph_break []
2025-12-04T09:45:16.5885269Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5885308Z Autotune Choices Stats:
2025-12-04T09:45:16.5886079Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.5886208Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5886323Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5886483Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5887106Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5887734Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5888348Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5888963Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5889568Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5890197Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5890839Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5891449Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5892086Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5892696Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5892828Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.5892868Z Autotune Choices Stats:
2025-12-04T09:45:16.5893633Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.5893852Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5894033Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5894332Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5894973Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5895605Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5896252Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5896880Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5897514Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5898142Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5898795Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5899421Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5900053Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5900735Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5900865Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.5900939Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5900984Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5901021Z unimplemented []
2025-12-04T09:45:16.5901082Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5901183Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5901763Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5901802Z graph_break []
2025-12-04T09:45:16.5901874Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5901916Z Autotune Choices Stats:
2025-12-04T09:45:16.5902661Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:16.5902820Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5902934Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5903097Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5903718Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5904328Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5904959Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5905566Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5906194Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5906823Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5907459Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5908065Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5908688Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5909312Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5909443Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:16.5909482Z Autotune Choices Stats:
2025-12-04T09:45:16.5910260Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:16.5910524Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5910691Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5910985Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5911636Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5912261Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5912888Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5913540Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5914171Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5914804Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5915426Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5916076Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5916707Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5917339Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5917478Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:16.5917551Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5917593Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5917646Z unimplemented []
2025-12-04T09:45:16.5917708Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5917810Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5918380Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5918418Z graph_break []
2025-12-04T09:45:16.5918492Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5918533Z Autotune Choices Stats:
2025-12-04T09:45:16.5919299Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.5919438Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5919552Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5919714Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5920343Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5921007Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5921614Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5922249Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5922856Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5923466Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5924096Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5924726Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5925329Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5925934Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5926073Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:16.5926114Z Autotune Choices Stats:
2025-12-04T09:45:16.5926890Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.5927108Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5927275Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5927553Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5928186Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5928834Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5929459Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5930082Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5930772Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5931400Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5932028Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5932656Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5933305Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5933930Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5934059Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:16.5934132Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5934196Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5934233Z unimplemented []
2025-12-04T09:45:16.5934294Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5934394Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5934987Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5935025Z graph_break []
2025-12-04T09:45:16.5935099Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5935139Z Autotune Choices Stats:
2025-12-04T09:45:16.5935888Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.5936018Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5936131Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5936294Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5936924Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5937547Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5938155Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5938764Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5939394Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5940006Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5940640Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5941255Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5941896Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5942502Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5942632Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:16.5942671Z Autotune Choices Stats:
2025-12-04T09:45:16.5943451Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.5943673Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5943840Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5944122Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5944759Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5945389Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5946037Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5946671Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5947321Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5947971Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5948602Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5949242Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5949886Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5950581Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5950711Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:16.5950784Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5950826Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5950866Z unimplemented []
2025-12-04T09:45:16.5950926Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5951028Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5951600Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5951651Z graph_break []
2025-12-04T09:45:16.5951724Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5951765Z Autotune Choices Stats:
2025-12-04T09:45:16.5952525Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:16.5952653Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5952770Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5952930Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5953544Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5954176Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5954781Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5955390Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5956003Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5956617Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5957229Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5957839Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5958456Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5959142Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5959272Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:16.5959314Z Autotune Choices Stats:
2025-12-04T09:45:16.5960075Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.5960304Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5960516Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5960796Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5961430Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5962074Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5962699Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5963349Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5963982Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5964631Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5965276Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5965906Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5966552Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5967204Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5967334Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:16.5967407Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5967452Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5967489Z unimplemented []
2025-12-04T09:45:16.5967551Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5967652Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5968233Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.5968272Z graph_break []
2025-12-04T09:45:16.5968347Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5968398Z Autotune Choices Stats:
2025-12-04T09:45:16.5969162Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:16.5969294Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5969409Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5969572Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5970188Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5970816Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5971452Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5972061Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5972672Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5973306Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5973911Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5974523Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5975128Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5975756Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5975885Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:16.5975926Z Autotune Choices Stats:
2025-12-04T09:45:16.5976692Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.5976912Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5977078Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5977372Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5978020Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5978648Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5979276Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5979927Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5980586Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5981220Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5981860Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5982509Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5983137Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5983760Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5983903Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:16.5983977Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.5984019Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.5984057Z unimplemented []
2025-12-04T09:45:16.5984117Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.5984233Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.5984810Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.5984849Z graph_break []
2025-12-04T09:45:16.5984920Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.5984962Z Autotune Choices Stats:
2025-12-04T09:45:16.5985707Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:16.5985847Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5985973Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5986134Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5986753Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5987362Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5987965Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5988598Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5989207Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5989834Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5990507Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5991118Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5991730Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5992339Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5992481Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:16.5992522Z Autotune Choices Stats:
2025-12-04T09:45:16.5993297Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:16.5993517Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.5993685Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.5993965Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.5994626Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5995255Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5995884Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5996509Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5997158Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5997785Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5998413Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.5999072Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.5999700Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6000329Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6000494Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:16.6000569Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6000613Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6000649Z unimplemented []
2025-12-04T09:45:16.6000725Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6000824Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6001415Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6001453Z graph_break []
2025-12-04T09:45:16.6001526Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6001566Z Autotune Choices Stats:
2025-12-04T09:45:16.6002320Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.6002447Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6002561Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6002736Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6003365Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6003973Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6004586Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6005188Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6005811Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6006420Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6007050Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6007676Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6008283Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6008893Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6009022Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:16.6009062Z Autotune Choices Stats:
2025-12-04T09:45:16.6009827Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.6010065Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6010231Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6010555Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6011193Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6011850Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6012477Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6013111Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6013744Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6014407Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6015034Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6015667Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6016317Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6016943Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6017074Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:16.6017169Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.6017216Z Traceback (most recent call last):
2025-12-04T09:45:16.6017372Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.6017412Z     self.assertTrue(
2025-12-04T09:45:16.6017519Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.6017569Z     raise self.failureException(msg)
2025-12-04T09:45:16.6017698Z AssertionError: False is not true : Log file /tmp/tmpun190rtr/flex_attention_configs.json was not created
2025-12-04T09:45:16.6017701Z 
2025-12-04T09:45:16.6017789Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.6017957Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.6017960Z 
2025-12-04T09:45:16.6018050Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.6018127Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6018169Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6018207Z unimplemented []
2025-12-04T09:45:16.6018277Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6018855Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.6018957Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6018994Z graph_break []
2025-12-04T09:45:16.6019068Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6019562Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.6019626Z   current_size = base.storage().size()
2025-12-04T09:45:16.6019668Z Autotune Choices Stats:
2025-12-04T09:45:16.6020469Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.6020599Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6020715Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6020880Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6021497Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6022102Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6022732Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6023334Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6023933Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6024562Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6025173Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6025778Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6026380Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6027007Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6027139Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.6027180Z Autotune Choices Stats:
2025-12-04T09:45:16.6027937Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.6028157Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6028333Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6028610Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6029251Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6029878Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6030554Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6031216Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6031840Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6032469Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6033115Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6033746Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6034374Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6034996Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6035136Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.6035212Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6035254Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6035292Z unimplemented []
2025-12-04T09:45:16.6035363Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6035463Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6036041Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6036083Z graph_break []
2025-12-04T09:45:16.6036156Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6036198Z Autotune Choices Stats:
2025-12-04T09:45:16.6036935Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.6037075Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6037204Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6037365Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6037979Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6038584Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6039187Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6039810Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6040442Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6041049Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6041683Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6042290Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6042902Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6043503Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6043646Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.6043690Z Autotune Choices Stats:
2025-12-04T09:45:16.6044459Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.6044683Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6044851Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6045133Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6045796Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6046423Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6047068Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6047711Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6048361Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6048990Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6049615Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6050261Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6050920Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6051547Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6051674Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.6051751Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6051793Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6051831Z unimplemented []
2025-12-04T09:45:16.6051909Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6052009Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6052597Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6052634Z graph_break []
2025-12-04T09:45:16.6052708Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6052749Z Autotune Choices Stats:
2025-12-04T09:45:16.6053493Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.6053622Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6053738Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6053911Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6054540Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6055144Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6055767Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6056386Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6057015Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6057629Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6058241Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6058867Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6059475Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6060077Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6060212Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.6060252Z Autotune Choices Stats:
2025-12-04T09:45:16.6061057Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.6061306Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6061474Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6061756Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6062398Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6063064Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6063689Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6064317Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6064947Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6065597Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6066226Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6066867Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6067517Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6068139Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6068269Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.6068341Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6068385Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6068422Z unimplemented []
2025-12-04T09:45:16.6068483Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6068584Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6069160Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6069209Z graph_break []
2025-12-04T09:45:16.6069281Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6069324Z Autotune Choices Stats:
2025-12-04T09:45:16.6070078Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.6070208Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6070323Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6070533Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6071149Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6071786Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6072396Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6073005Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6073617Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6074243Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6074852Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6075458Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6076085Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6076686Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6076819Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.6076861Z Autotune Choices Stats:
2025-12-04T09:45:16.6077635Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.6077867Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6078036Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6078329Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6078961Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6079590Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6080234Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6080899Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6081528Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6082154Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6082808Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6083440Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6084083Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6084730Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6084858Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.6084933Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6084976Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6085014Z unimplemented []
2025-12-04T09:45:16.6085075Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6085177Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6085757Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6085794Z graph_break []
2025-12-04T09:45:16.6085867Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6085907Z Autotune Choices Stats:
2025-12-04T09:45:16.6086656Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.6086804Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6086920Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6087082Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6087694Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6088298Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6088928Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6089531Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6090137Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6090799Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6091431Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6092037Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6092637Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6093269Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6093398Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.6093440Z Autotune Choices Stats:
2025-12-04T09:45:16.6094199Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.6094418Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6094585Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6097588Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6098234Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6098858Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6099484Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6100126Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6100793Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6101420Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6102040Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6102691Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6103317Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6103943Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6104085Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.6104158Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6104212Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6104250Z unimplemented []
2025-12-04T09:45:16.6104311Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6104410Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6104992Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6105031Z graph_break []
2025-12-04T09:45:16.6105104Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6105148Z Autotune Choices Stats:
2025-12-04T09:45:16.6105897Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.6106036Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6106150Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6106314Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6106938Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6107538Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6108146Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6108772Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6109374Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6109980Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6110626Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6111253Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6111860Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6112467Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6112609Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.6112649Z Autotune Choices Stats:
2025-12-04T09:45:16.6113443Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.6113664Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6113831Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6114114Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6114752Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6115403Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6116030Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6116663Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6117308Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6117934Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6118559Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6119184Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6119829Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6120494Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6120624Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.6120700Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6120756Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6120794Z unimplemented []
2025-12-04T09:45:16.6120854Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6120958Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6121550Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6121588Z graph_break []
2025-12-04T09:45:16.6121661Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6121703Z Autotune Choices Stats:
2025-12-04T09:45:16.6122448Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.6122577Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6122690Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6122852Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6123490Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6124089Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6124698Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6125304Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6125929Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6126531Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6127139Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6127745Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6128367Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6128973Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6129103Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.6129161Z Autotune Choices Stats:
2025-12-04T09:45:16.6129939Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.6130158Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6130325Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6130632Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6131265Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6131892Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6132541Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6133165Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6133795Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6134448Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6135073Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6135703Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6136330Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6136970Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6137101Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.6137174Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6137218Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6137256Z unimplemented []
2025-12-04T09:45:16.6137316Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6137417Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6137999Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6138047Z graph_break []
2025-12-04T09:45:16.6138122Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6138162Z Autotune Choices Stats:
2025-12-04T09:45:16.6138925Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.6139054Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6139167Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6139329Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6139946Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6140619Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6141228Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6141834Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6142456Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6143075Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6143684Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6144287Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6144904Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6145524Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6145655Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.6145696Z Autotune Choices Stats:
2025-12-04T09:45:16.6146455Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.6146687Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6146862Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6147144Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6147780Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6148406Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6149031Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6149677Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6150309Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6150969Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6151621Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6152251Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6152876Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6153521Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6153651Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.6153726Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6153770Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6153808Z unimplemented []
2025-12-04T09:45:16.6153867Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6153970Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6154550Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6154588Z graph_break []
2025-12-04T09:45:16.6154661Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6154712Z Autotune Choices Stats:
2025-12-04T09:45:16.6155463Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.6155590Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6155706Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6155870Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6156489Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6157098Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6157724Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6158329Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6158942Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6159552Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6160167Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6160811Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6161417Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6162058Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6162186Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.6162229Z Autotune Choices Stats:
2025-12-04T09:45:16.6162992Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.6163213Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6163382Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6163673Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6164318Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6164943Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6165581Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6166230Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6166856Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6167482Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6168118Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6168772Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6169400Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6170025Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6170166Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.6170239Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6170283Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6170320Z unimplemented []
2025-12-04T09:45:16.6170382Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6170530Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6171109Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6171147Z graph_break []
2025-12-04T09:45:16.6171221Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6171262Z Autotune Choices Stats:
2025-12-04T09:45:16.6172010Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.6172151Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6172266Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6172439Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6173048Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6173649Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6174257Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6174886Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6175487Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6176093Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6176728Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6177334Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6177937Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6178539Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6178682Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.6178723Z Autotune Choices Stats:
2025-12-04T09:45:16.6179502Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.6179722Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6179888Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6180167Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6180843Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6181495Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6182124Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6182770Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6183447Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6184074Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6184698Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6185348Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6185979Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6186603Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6186735Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.6186810Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6186854Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6186892Z unimplemented []
2025-12-04T09:45:16.6186951Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6187066Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6187654Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6187692Z graph_break []
2025-12-04T09:45:16.6187765Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6187806Z Autotune Choices Stats:
2025-12-04T09:45:16.6188549Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.6188678Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6188793Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6188965Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6189592Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6190204Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6190845Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6191444Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6192085Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6192696Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6193306Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6193956Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6194566Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6195184Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6195312Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.6195353Z Autotune Choices Stats:
2025-12-04T09:45:16.6196133Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.6196378Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6196545Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6196834Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6197472Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6198121Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6198749Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6199395Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6200021Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6200704Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6201327Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6201956Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6202625Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6203254Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6203386Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.6203460Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6203506Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6203544Z unimplemented []
2025-12-04T09:45:16.6203604Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6203705Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6204281Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6204340Z graph_break []
2025-12-04T09:45:16.6204414Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6204454Z Autotune Choices Stats:
2025-12-04T09:45:16.6205212Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.6205341Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6205455Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6205617Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6206231Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6206863Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6207469Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6208078Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6208680Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6209316Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6209922Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6210571Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6211218Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6211821Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6211952Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.6211993Z Autotune Choices Stats:
2025-12-04T09:45:16.6212756Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.6212993Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6213158Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6213461Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6214092Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6214725Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6215370Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6215997Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6216620Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6217249Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6217889Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6218518Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6219147Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6219798Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6219929Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.6220003Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6220047Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6220085Z unimplemented []
2025-12-04T09:45:16.6220146Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6220249Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6220862Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6220900Z graph_break []
2025-12-04T09:45:16.6220972Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6221014Z Autotune Choices Stats:
2025-12-04T09:45:16.6221760Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.6221933Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6222048Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6222208Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6222820Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6223423Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6224074Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6224684Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6225291Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6225898Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6226538Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6227139Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6227739Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6228374Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6228502Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.6228543Z Autotune Choices Stats:
2025-12-04T09:45:16.6229306Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.6229529Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6229695Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6229987Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6230684Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6231312Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6231938Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6232607Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6233235Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6233858Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6234476Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6235149Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6235775Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6236399Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6236540Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.6236613Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6236655Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6236706Z unimplemented []
2025-12-04T09:45:16.6236767Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6236869Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6237446Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6237485Z graph_break []
2025-12-04T09:45:16.6237559Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6237599Z Autotune Choices Stats:
2025-12-04T09:45:16.6238342Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.6238486Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6238598Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6238758Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6239384Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6239988Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6240651Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6241307Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6241910Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6242510Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6246518Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6247223Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6247864Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6248470Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6248616Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.6248659Z Autotune Choices Stats:
2025-12-04T09:45:16.6249435Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.6249661Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6249839Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6250128Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6250803Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6251459Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6252085Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6252726Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6253383Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6254012Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6254645Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6255298Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6255942Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6256570Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6256703Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.6256784Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6256840Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6256879Z unimplemented []
2025-12-04T09:45:16.6256943Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6257045Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6257635Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6257675Z graph_break []
2025-12-04T09:45:16.6257750Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6257792Z Autotune Choices Stats:
2025-12-04T09:45:16.6258545Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.6258678Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6258796Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6258961Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6259588Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6260199Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6260844Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6261451Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6262097Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6262702Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6263313Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6263936Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6264565Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6265171Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6265303Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.6265345Z Autotune Choices Stats:
2025-12-04T09:45:16.6266129Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.6266353Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6266522Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6266806Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6267449Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6268088Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6268730Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6269358Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6270007Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6270771Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6271395Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6272027Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6272662Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6273318Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6273448Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.6273524Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6273567Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6273606Z unimplemented []
2025-12-04T09:45:16.6273668Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6273770Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6274356Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6274409Z graph_break []
2025-12-04T09:45:16.6274484Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6274525Z Autotune Choices Stats:
2025-12-04T09:45:16.6275282Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.6275411Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6275527Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6275690Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6276309Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6276940Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6277547Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6278155Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6278764Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6279392Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6280003Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6280640Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6281252Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6281880Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6282012Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.6282054Z Autotune Choices Stats:
2025-12-04T09:45:16.6282820Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.6283052Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6283232Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6283512Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6284142Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6284779Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6285405Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6286059Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6286692Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6287322Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6287970Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6288599Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6289230Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6289869Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6290012Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.6290087Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6290132Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6290170Z unimplemented []
2025-12-04T09:45:16.6290233Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6290334Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6290946Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6290985Z graph_break []
2025-12-04T09:45:16.6291058Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6291113Z Autotune Choices Stats:
2025-12-04T09:45:16.6291868Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.6292000Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6292114Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6292276Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6292887Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6293499Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6294147Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6294753Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6295360Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6295976Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6296599Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6297206Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6297817Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6298428Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6298568Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.6298610Z Autotune Choices Stats:
2025-12-04T09:45:16.6299372Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.6299592Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6299758Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6300049Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6300739Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6301368Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6301995Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6302630Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6303289Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6303921Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6304561Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6305217Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6305840Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6306469Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6306612Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.6306687Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6306729Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6306768Z unimplemented []
2025-12-04T09:45:16.6306828Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6306928Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6307516Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6307555Z graph_break []
2025-12-04T09:45:16.6307630Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6307670Z Autotune Choices Stats:
2025-12-04T09:45:16.6308420Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.6308557Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6308672Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6308850Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6309461Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6310067Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6310711Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6311344Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6311949Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6312571Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6313206Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6313813Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6314424Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6315043Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6315183Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.6315228Z Autotune Choices Stats:
2025-12-04T09:45:16.6316003Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.6316223Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6316394Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6316678Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6317313Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6317961Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6318593Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6319212Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6319858Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6320528Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6321151Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6321796Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6322438Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6323062Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6323195Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.6323272Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6323316Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6323353Z unimplemented []
2025-12-04T09:45:16.6323416Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6323517Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6324120Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6324157Z graph_break []
2025-12-04T09:45:16.6324232Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6324272Z Autotune Choices Stats:
2025-12-04T09:45:16.6325023Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.6325153Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6325268Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6325445Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6326070Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6326674Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6327276Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6327876Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6328503Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6329107Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6329717Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6330345Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6330995Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6331607Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6331739Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.6331780Z Autotune Choices Stats:
2025-12-04T09:45:16.6332556Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.6332816Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6332983Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6333263Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6333902Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6334544Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6335181Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6335807Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6336439Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6337088Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6337714Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6338348Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6338996Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6339625Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6339755Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.6339830Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6339872Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6339910Z unimplemented []
2025-12-04T09:45:16.6339971Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6340069Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6341063Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6341119Z graph_break []
2025-12-04T09:45:16.6341192Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6341235Z Autotune Choices Stats:
2025-12-04T09:45:16.6341998Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.6342125Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6342241Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6342403Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6343021Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6343655Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6344260Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6344860Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6345463Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6346090Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6346697Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6347306Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6347932Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6348533Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6348664Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.6348706Z Autotune Choices Stats:
2025-12-04T09:45:16.6349582Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.6349802Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6349982Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6350278Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6350943Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6351574Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6352224Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6352852Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6353484Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6354114Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6354773Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6355405Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6356040Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6356686Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6356816Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.6356891Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6356935Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6356973Z unimplemented []
2025-12-04T09:45:16.6357035Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6357135Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6357716Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6357754Z graph_break []
2025-12-04T09:45:16.6357828Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6357868Z Autotune Choices Stats:
2025-12-04T09:45:16.6358615Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.6358771Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6358885Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6359048Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6359666Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6360266Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6360923Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6361529Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6362138Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6362760Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6363394Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6364005Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6364611Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6365236Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6365367Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.6365407Z Autotune Choices Stats:
2025-12-04T09:45:16.6366170Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.6366391Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6366557Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6366835Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6367490Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6368122Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6368749Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6369394Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6370024Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6370698Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6371323Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6371977Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6372610Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6373253Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6373395Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.6373470Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6373512Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6373551Z unimplemented []
2025-12-04T09:45:16.6373630Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6373731Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6374307Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6374351Z graph_break []
2025-12-04T09:45:16.6374424Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6374466Z Autotune Choices Stats:
2025-12-04T09:45:16.6375221Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.6375360Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6375475Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6375637Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6376265Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6376871Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6377480Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6378104Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6378712Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6379324Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6379934Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6380588Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6381196Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6381817Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6381958Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.6382001Z Autotune Choices Stats:
2025-12-04T09:45:16.6382780Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.6383000Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6383169Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6383449Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6384084Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6384729Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6385355Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6385980Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6386631Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6387262Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6387904Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6388538Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6389188Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6389816Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6389946Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.6390020Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6390082Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6390119Z unimplemented []
2025-12-04T09:45:16.6390179Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6390277Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6390915Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6390953Z graph_break []
2025-12-04T09:45:16.6391027Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6391068Z Autotune Choices Stats:
2025-12-04T09:45:16.6391814Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.6391943Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6392057Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6392220Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6392843Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6393475Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6394084Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6394696Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6395327Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6395934Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6396548Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6397161Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6397788Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6398397Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6398527Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.6398567Z Autotune Choices Stats:
2025-12-04T09:45:16.6399343Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.6399575Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6399741Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6400023Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6400703Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6401328Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6401987Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6402617Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6403249Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6403905Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6404533Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6405165Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6405795Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6406450Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6406580Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.6406654Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6406695Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6406733Z unimplemented []
2025-12-04T09:45:16.6406794Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6406893Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6407464Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6407513Z graph_break []
2025-12-04T09:45:16.6407586Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6407628Z Autotune Choices Stats:
2025-12-04T09:45:16.6408385Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.6408516Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6408630Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6408792Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6409413Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6410030Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6410681Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6411290Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6411912Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6412541Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6413158Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6413767Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6414394Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6415022Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6415153Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.6415195Z Autotune Choices Stats:
2025-12-04T09:45:16.6415960Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.6416188Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6416368Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6416648Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6417285Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6417912Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6418545Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6419191Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6419821Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6420487Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6421142Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6421774Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6422399Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6423046Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6423187Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.6423263Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6423308Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6423347Z unimplemented []
2025-12-04T09:45:16.6423407Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6423508Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6424085Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6424121Z graph_break []
2025-12-04T09:45:16.6424195Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6424245Z Autotune Choices Stats:
2025-12-04T09:45:16.6425002Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.6425132Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6425246Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6425409Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6426032Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6426640Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6427267Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6427874Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6428485Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6429101Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6429739Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6430350Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6431008Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6431630Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6431771Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.6431812Z Autotune Choices Stats:
2025-12-04T09:45:16.6432579Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.6432800Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6432968Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6433263Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6433914Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6434544Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6435171Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6435815Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6436456Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6437092Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6437731Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6438389Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6439021Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6439650Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6439794Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.6439867Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6439910Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6439948Z unimplemented []
2025-12-04T09:45:16.6440009Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6440119Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6440733Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6440772Z graph_break []
2025-12-04T09:45:16.6440846Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6440889Z Autotune Choices Stats:
2025-12-04T09:45:16.6441641Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.6441785Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6441900Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6442077Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6442698Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6443308Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6443916Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6444548Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6445154Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6445765Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6446395Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6447001Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6448534Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6449156Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6449312Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.6449355Z Autotune Choices Stats:
2025-12-04T09:45:16.6450142Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.6450364Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6450558Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6450840Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6451472Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6452122Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6452752Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6453410Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6454075Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6454712Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6455346Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6455986Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6456613Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6457256Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6457385Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.6457461Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6457503Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6457542Z unimplemented []
2025-12-04T09:45:16.6457602Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6457715Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6458305Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6458344Z graph_break []
2025-12-04T09:45:16.6458419Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6458459Z Autotune Choices Stats:
2025-12-04T09:45:16.6459214Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.6459343Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6459459Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6459621Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6460249Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6460889Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6461516Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6462129Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6462761Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6463370Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6463982Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6464605Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6465218Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6465848Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6465978Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.6466019Z Autotune Choices Stats:
2025-12-04T09:45:16.6466791Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.6467034Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6467202Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6467486Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6468116Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6468754Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6469386Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6470025Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6470698Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6471362Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6471996Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6472640Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6473301Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6473928Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6474072Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.6474145Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6474190Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6474227Z unimplemented []
2025-12-04T09:45:16.6474289Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6474390Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6474970Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6475022Z graph_break []
2025-12-04T09:45:16.6475095Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6475136Z Autotune Choices Stats:
2025-12-04T09:45:16.6475898Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.6476030Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6476144Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6476307Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6476924Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6477542Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6478146Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6478765Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6479372Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6479998Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6480637Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6481244Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6481865Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6482474Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6482619Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.6482661Z Autotune Choices Stats:
2025-12-04T09:45:16.6483425Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.6483658Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6483825Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6484128Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6484757Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6485390Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6486020Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6486646Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6487298Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6487926Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6488579Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6489204Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6489837Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6490508Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6490637Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.6490711Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6490754Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6490792Z unimplemented []
2025-12-04T09:45:16.6490853Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6490968Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6491548Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6491586Z graph_break []
2025-12-04T09:45:16.6491660Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6491700Z Autotune Choices Stats:
2025-12-04T09:45:16.6492455Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.6492608Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6492723Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6492884Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6493504Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6494132Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6494753Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6495361Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6495976Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6496579Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6497208Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6497815Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6498440Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6499057Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6499186Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.6499227Z Autotune Choices Stats:
2025-12-04T09:45:16.6499996Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.6500228Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6500395Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6500709Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6501368Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6501997Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6502632Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6503279Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6503916Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6504562Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6505198Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6505849Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6506479Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6507112Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6507243Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.6507316Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6507372Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6507410Z unimplemented []
2025-12-04T09:45:16.6507471Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6507571Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6508149Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6508200Z graph_break []
2025-12-04T09:45:16.6508274Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6508316Z Autotune Choices Stats:
2025-12-04T09:45:16.6509057Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:16.6509197Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6509312Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6509475Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6510105Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6510752Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6511361Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6512001Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6512610Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6513238Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6513851Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6514509Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6515119Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6515727Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6515857Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:16.6515898Z Autotune Choices Stats:
2025-12-04T09:45:16.6516673Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:16.6516903Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6517071Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6517353Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6517989Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6518650Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6519280Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6519914Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6520601Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6521234Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6521900Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6522534Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6523198Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6523829Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6523958Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:16.6524033Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6524074Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6524112Z unimplemented []
2025-12-04T09:45:16.6524174Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6524276Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6524868Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6524907Z graph_break []
2025-12-04T09:45:16.6524979Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6525020Z Autotune Choices Stats:
2025-12-04T09:45:16.6525774Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.6525917Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6526033Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6526195Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6526835Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6527447Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6528074Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6528682Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6529302Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6529913Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6530562Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6531188Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6531823Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6532435Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6532566Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:16.6532607Z Autotune Choices Stats:
2025-12-04T09:45:16.6533396Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.6533617Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6533782Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6534074Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6534715Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6535344Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6536000Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6536634Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6537259Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6537910Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6538536Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6539183Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6539822Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6540505Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6540636Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:16.6540713Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6540757Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6540795Z unimplemented []
2025-12-04T09:45:16.6540855Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6540957Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6541533Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6541571Z graph_break []
2025-12-04T09:45:16.6541646Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6541698Z Autotune Choices Stats:
2025-12-04T09:45:16.6542439Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.6542582Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6542697Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6542861Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6543475Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6544109Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6544717Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6545321Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6545937Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6546548Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6547162Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6547783Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6548414Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6549020Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6549151Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:16.6549191Z Autotune Choices Stats:
2025-12-04T09:45:16.6549956Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.6550176Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6550351Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6550672Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6551300Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6551959Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6552604Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6553242Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6553879Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6554527Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6555190Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6555817Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6556459Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6557103Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6557232Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:16.6557307Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6557348Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6557388Z unimplemented []
2025-12-04T09:45:16.6557450Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6557550Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6558123Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6558160Z graph_break []
2025-12-04T09:45:16.6558233Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6558274Z Autotune Choices Stats:
2025-12-04T09:45:16.6559043Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:16.6559172Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6559288Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6559448Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6560078Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6560723Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6561358Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6561968Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6562578Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6563199Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6563814Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6564435Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6565043Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6565673Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6565802Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:16.6565844Z Autotune Choices Stats:
2025-12-04T09:45:16.6566614Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.6566835Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6567003Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6567285Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6567941Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6568568Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6569208Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6569854Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6570505Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6571139Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6571780Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6572415Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6573060Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6573686Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6573829Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:16.6573906Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6573949Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6573985Z unimplemented []
2025-12-04T09:45:16.6574057Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6574157Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6574739Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6574777Z graph_break []
2025-12-04T09:45:16.6574853Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6574892Z Autotune Choices Stats:
2025-12-04T09:45:16.6575647Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:16.6575775Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6575900Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6576063Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6576679Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6577303Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6577913Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6578544Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6579151Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6579767Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6580380Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6581025Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6581657Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6582263Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6582404Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:16.6582444Z Autotune Choices Stats:
2025-12-04T09:45:16.6583228Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.6583449Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6583618Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6583895Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6584540Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6585169Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6585807Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6586441Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6587098Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6587732Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6588357Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6589004Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6589640Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6590279Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6590692Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:16.6590767Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6590809Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6590865Z unimplemented []
2025-12-04T09:45:16.6590926Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6591027Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6591616Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6591654Z graph_break []
2025-12-04T09:45:16.6591728Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6591770Z Autotune Choices Stats:
2025-12-04T09:45:16.6592511Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:16.6592641Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6592758Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6592920Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6593551Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6594160Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6594781Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6595405Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6596024Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6596631Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6597246Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6597874Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6598481Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6599099Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6599228Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:16.6599270Z Autotune Choices Stats:
2025-12-04T09:45:16.6600057Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:16.6600286Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6600485Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6600769Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6601408Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6602072Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6602699Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6603340Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6603971Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6604631Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6605259Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6605908Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6606547Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6607173Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6607311Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:16.6607387Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6607429Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6607467Z unimplemented []
2025-12-04T09:45:16.6607527Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6607628Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6608210Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6608261Z graph_break []
2025-12-04T09:45:16.6608336Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6608375Z Autotune Choices Stats:
2025-12-04T09:45:16.6609138Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.6609268Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6609383Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6609548Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6610168Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6610849Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6611458Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6612089Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6612695Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6613341Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6613954Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6614567Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6615189Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6615793Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6615940Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:16.6615981Z Autotune Choices Stats:
2025-12-04T09:45:16.6616745Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.6616976Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6617153Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6617436Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6618079Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6618707Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6619346Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6619975Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6620643Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6621278Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6621945Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6622579Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6623210Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6623859Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6623990Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:16.6624064Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6624108Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6624164Z unimplemented []
2025-12-04T09:45:16.6624225Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6624324Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6624896Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6624933Z graph_break []
2025-12-04T09:45:16.6625004Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6625046Z Autotune Choices Stats:
2025-12-04T09:45:16.6625806Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.6625934Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6626049Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6626213Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6626838Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6627455Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6628075Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6628692Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6629312Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6629922Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6630598Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6631206Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6631816Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6632442Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6632569Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:16.6632611Z Autotune Choices Stats:
2025-12-04T09:45:16.6633379Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.6633611Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6633777Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6634071Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6634719Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6635344Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6635972Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6636609Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6637244Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6637891Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6638523Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6639171Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6639804Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6640465Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6640596Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:16.6640704Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.6640753Z Traceback (most recent call last):
2025-12-04T09:45:16.6640911Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.6640953Z     self.assertTrue(
2025-12-04T09:45:16.6641061Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.6641110Z     raise self.failureException(msg)
2025-12-04T09:45:16.6641240Z AssertionError: False is not true : Log file /tmp/tmpwzy0l12r/flex_attention_configs.json was not created
2025-12-04T09:45:16.6641258Z 
2025-12-04T09:45:16.6641335Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.6641501Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.6641504Z 
2025-12-04T09:45:16.6641598Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.6641673Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6641719Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6641757Z unimplemented []
2025-12-04T09:45:16.6641821Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6642409Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.6642526Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6642562Z graph_break []
2025-12-04T09:45:16.6642636Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6643144Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.6643195Z   current_size = base.storage().size()
2025-12-04T09:45:16.6643240Z Autotune Choices Stats:
2025-12-04T09:45:16.6643983Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.6644114Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6644228Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6644394Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6645023Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6645627Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6646243Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6646850Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6647480Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6648083Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6648690Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6649310Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6649917Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6650592Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6650723Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.6650764Z Autotune Choices Stats:
2025-12-04T09:45:16.6651559Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.6651780Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6651947Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6652230Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6652881Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6653544Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6654168Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6654806Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6655448Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6656100Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6656721Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6657355Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6658015Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6658638Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6658780Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.6658857Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6658899Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6658938Z unimplemented []
2025-12-04T09:45:16.6658998Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6659098Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6659671Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6659719Z graph_break []
2025-12-04T09:45:16.6659795Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6659834Z Autotune Choices Stats:
2025-12-04T09:45:16.6660626Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.6660757Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6660874Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6661036Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6661655Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6662275Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6662884Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6663501Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6664108Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6664744Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6665355Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6665966Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6666592Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6667199Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6667342Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.6667382Z Autotune Choices Stats:
2025-12-04T09:45:16.6668154Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.6668384Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6668566Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6668846Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6669497Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6670121Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6670798Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6671423Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6672064Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6672689Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6673334Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6673967Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6674596Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6675249Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6675379Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.6675453Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6675499Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6675549Z unimplemented []
2025-12-04T09:45:16.6675610Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6675710Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6676287Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6676326Z graph_break []
2025-12-04T09:45:16.6676399Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6676440Z Autotune Choices Stats:
2025-12-04T09:45:16.6677204Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.6677333Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6677447Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6677612Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6678233Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6678840Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6679476Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6680080Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6680745Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6681349Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6681976Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6682586Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6683194Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6683813Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6683950Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.6683989Z Autotune Choices Stats:
2025-12-04T09:45:16.6684773Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.6685008Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6685174Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6685468Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6686121Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6686746Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6687373Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6688009Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6688639Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6689282Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6689912Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6690591Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6691222Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6691864Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6691994Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.6692080Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6692123Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6692161Z unimplemented []
2025-12-04T09:45:16.6692222Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6692322Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6692904Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6692957Z graph_break []
2025-12-04T09:45:16.6693032Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6693074Z Autotune Choices Stats:
2025-12-04T09:45:16.6693829Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.6693969Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6694086Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6694257Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6694869Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6695476Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6696084Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6696709Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6697310Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6697926Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6698552Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6699175Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6699778Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6700382Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6700552Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.6700608Z Autotune Choices Stats:
2025-12-04T09:45:16.6701372Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.6701606Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6701775Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6702051Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6702688Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6703336Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6703957Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6704582Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6705224Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6705852Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6706490Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6707137Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6707781Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6708408Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6708538Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.6708611Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6708656Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6708695Z unimplemented []
2025-12-04T09:45:16.6708757Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6708856Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6709439Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6709477Z graph_break []
2025-12-04T09:45:16.6709553Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6709592Z Autotune Choices Stats:
2025-12-04T09:45:16.6710354Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.6710515Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6710629Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6710806Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6711447Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6712048Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6712660Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6713276Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6713899Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6714509Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6715133Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6715766Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6716371Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6716981Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6717230Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.6717271Z Autotune Choices Stats:
2025-12-04T09:45:16.6718046Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.6718268Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6718441Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6718731Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6719371Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6719989Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6720661Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6721289Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6721938Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6722603Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6723228Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6723875Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6724532Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6725160Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6725292Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.6725368Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6725410Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6725450Z unimplemented []
2025-12-04T09:45:16.6725512Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6725613Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6726186Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6726227Z graph_break []
2025-12-04T09:45:16.6726311Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6726353Z Autotune Choices Stats:
2025-12-04T09:45:16.6727103Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.6727248Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6727364Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6727525Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6728150Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6728792Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6729400Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6730012Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6730703Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6731304Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6731926Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6732534Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6733172Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6733781Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6733914Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.6733958Z Autotune Choices Stats:
2025-12-04T09:45:16.6734725Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.6734946Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6735124Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6735402Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6736039Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6736680Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6737329Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6737948Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6738579Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6739209Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6739849Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6740509Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6741150Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6741805Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6741936Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.6742010Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6742054Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6742095Z unimplemented []
2025-12-04T09:45:16.6742157Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6742258Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6742839Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6742877Z graph_break []
2025-12-04T09:45:16.6742953Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6742993Z Autotune Choices Stats:
2025-12-04T09:45:16.6743757Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.6743890Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6744004Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6744168Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6744807Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6745411Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6746044Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6746650Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6747256Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6747876Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6748483Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6749104Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6749712Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6750341Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6750512Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.6750551Z Autotune Choices Stats:
2025-12-04T09:45:16.6751321Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.6751543Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6751708Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6751993Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6752643Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6753270Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6755892Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6756555Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6757189Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6757819Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6758449Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6759080Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6759725Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6760351Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6760529Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.6760610Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6760653Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6760693Z unimplemented []
2025-12-04T09:45:16.6760771Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6760874Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6761449Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6761489Z graph_break []
2025-12-04T09:45:16.6761566Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6761608Z Autotune Choices Stats:
2025-12-04T09:45:16.6762364Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.6762495Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6762633Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6762794Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6763414Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6764036Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6764634Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6765264Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6765870Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6766479Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6767097Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6767714Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6768342Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6768956Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6769102Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.6769145Z Autotune Choices Stats:
2025-12-04T09:45:16.6769936Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.6770163Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6770335Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6770646Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6771298Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6771928Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6772569Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6773194Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6773856Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6774485Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6775130Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6775771Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6776398Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6777039Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6777168Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.6777244Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6777287Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6777339Z unimplemented []
2025-12-04T09:45:16.6777400Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6777502Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6778095Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6778132Z graph_break []
2025-12-04T09:45:16.6778207Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6778247Z Autotune Choices Stats:
2025-12-04T09:45:16.6779008Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.6779136Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6779250Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6779412Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6780043Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6780681Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6781301Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6781911Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6782548Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6783146Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6783755Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6784370Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6784976Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6785584Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6785714Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.6785754Z Autotune Choices Stats:
2025-12-04T09:45:16.6786532Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.6786761Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6786927Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6787207Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6787840Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6788480Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6789119Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6789765Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6790438Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6791102Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6791754Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6792395Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6793061Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6793695Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6793837Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.6793911Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6793954Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6793992Z unimplemented []
2025-12-04T09:45:16.6794054Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6794154Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6794738Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6794793Z graph_break []
2025-12-04T09:45:16.6794868Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6794909Z Autotune Choices Stats:
2025-12-04T09:45:16.6795664Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.6795793Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6795909Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6796069Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6796685Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6797303Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6797910Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6798524Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6799130Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6799750Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6800357Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6800998Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6801623Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6802231Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6802372Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.6802414Z Autotune Choices Stats:
2025-12-04T09:45:16.6803170Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.6803402Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6803582Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6803860Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6804517Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6805144Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6805773Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6806401Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6807044Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6807672Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6808320Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6808963Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6809609Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6810253Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6810381Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.6810499Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6810540Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6810582Z unimplemented []
2025-12-04T09:45:16.6810663Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6810765Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6811348Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6811387Z graph_break []
2025-12-04T09:45:16.6811462Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6811502Z Autotune Choices Stats:
2025-12-04T09:45:16.6812260Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.6812404Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6812518Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6812679Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6813307Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6813905Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6814533Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6815152Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6815773Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6816376Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6817016Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6817627Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6818236Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6818871Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6819001Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.6819041Z Autotune Choices Stats:
2025-12-04T09:45:16.6819799Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.6820034Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6820199Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6820540Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6821197Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6821824Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6822471Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6823115Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6823746Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6824391Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6825025Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6825683Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6826308Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6826941Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6827074Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.6827163Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6827206Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6827243Z unimplemented []
2025-12-04T09:45:16.6827305Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6827405Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6827985Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6828035Z graph_break []
2025-12-04T09:45:16.6828108Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6828150Z Autotune Choices Stats:
2025-12-04T09:45:16.6828894Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.6829035Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6829151Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6829311Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6829940Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6830579Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6831186Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6831801Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6832404Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6833024Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6833637Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6834270Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6834880Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6835485Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6835616Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.6835657Z Autotune Choices Stats:
2025-12-04T09:45:16.6836428Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.6836661Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6836829Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6837109Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6837745Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6838391Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6839016Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6839658Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6840319Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6840976Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6841615Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6842246Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6842911Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6843539Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6843666Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.6843741Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6843783Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6843822Z unimplemented []
2025-12-04T09:45:16.6843883Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6843984Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6844585Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6844623Z graph_break []
2025-12-04T09:45:16.6844699Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6844739Z Autotune Choices Stats:
2025-12-04T09:45:16.6845487Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.6845624Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6845739Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6845904Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6846541Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6847150Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6847764Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6848368Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6848990Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6849597Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6850216Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6850869Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6851484Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6852090Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6852220Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.6852260Z Autotune Choices Stats:
2025-12-04T09:45:16.6853035Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.6853254Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6853421Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6853714Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6854363Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6854993Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6855642Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6856274Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6856904Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6857548Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6858175Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6858814Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6859458Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6860111Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6860243Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.6860318Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6860362Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6860400Z unimplemented []
2025-12-04T09:45:16.6860494Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6860595Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6861173Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6861213Z graph_break []
2025-12-04T09:45:16.6861286Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6861342Z Autotune Choices Stats:
2025-12-04T09:45:16.6862085Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.6862228Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6862344Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6862508Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6863124Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6863756Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6864362Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6864987Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6865593Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6866205Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6866816Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6867436Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6868062Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6868666Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6868797Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.6868837Z Autotune Choices Stats:
2025-12-04T09:45:16.6869602Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.6869822Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6869998Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6870277Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6870959Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6871597Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6872225Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6872857Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6873490Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6874121Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6874776Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6875404Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6876054Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6876703Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6876830Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.6876911Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6876953Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6876992Z unimplemented []
2025-12-04T09:45:16.6877053Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6877155Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6877731Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6877768Z graph_break []
2025-12-04T09:45:16.6877844Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6877883Z Autotune Choices Stats:
2025-12-04T09:45:16.6878640Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.6878768Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6878884Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6879042Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6879668Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6880289Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6880951Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6881554Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6882161Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6882796Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6883407Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6884029Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6884637Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6885268Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6885396Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.6885437Z Autotune Choices Stats:
2025-12-04T09:45:16.6886202Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.6886421Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6886587Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6886870Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6887514Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6888141Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6888775Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6889425Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6890055Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6890725Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6891366Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6891999Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6892629Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6893262Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6893403Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.6893478Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6893522Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6893560Z unimplemented []
2025-12-04T09:45:16.6893620Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6893733Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6894312Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6894351Z graph_break []
2025-12-04T09:45:16.6894426Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6894466Z Autotune Choices Stats:
2025-12-04T09:45:16.6895210Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.6895340Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6895464Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6895626Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6896252Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6896867Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6897467Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6898095Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6898701Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6899322Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6899938Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6900576Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6901203Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6901807Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6901948Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.6901988Z Autotune Choices Stats:
2025-12-04T09:45:16.6902764Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.6902985Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6903152Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6903434Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6904078Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6904697Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6905323Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6905955Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6906611Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6907239Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6907867Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6908506Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6909135Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6909774Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6909902Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.6909975Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6910017Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6910056Z unimplemented []
2025-12-04T09:45:16.6910126Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6910229Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6910857Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6910896Z graph_break []
2025-12-04T09:45:16.6910968Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6911009Z Autotune Choices Stats:
2025-12-04T09:45:16.6911762Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.6911890Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6912005Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6912166Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6912792Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6913398Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6914024Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6914627Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6915249Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6915854Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6916471Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6917092Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6917699Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6918317Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6918446Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.6918487Z Autotune Choices Stats:
2025-12-04T09:45:16.6919244Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.6919483Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6919650Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6919929Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6920596Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6921235Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6921861Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6922500Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6923135Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6923796Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6924421Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6925069Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6925710Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6926339Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6926478Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.6926552Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6926595Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6926633Z unimplemented []
2025-12-04T09:45:16.6926695Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6926796Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6927380Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6927428Z graph_break []
2025-12-04T09:45:16.6927502Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6927542Z Autotune Choices Stats:
2025-12-04T09:45:16.6928295Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.6928425Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6928538Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6928700Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6929316Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6929936Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6930577Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6931198Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6931804Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6932435Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6933039Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6933648Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6934288Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6934893Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6935034Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.6935074Z Autotune Choices Stats:
2025-12-04T09:45:16.6935836Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.6936067Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6936237Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6936523Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6937158Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6937779Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6938407Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6939038Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6939685Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6940313Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6940989Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6941623Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6942267Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6942905Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6943034Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.6943108Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6943150Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6943189Z unimplemented []
2025-12-04T09:45:16.6943251Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6943368Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6943947Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6943985Z graph_break []
2025-12-04T09:45:16.6944057Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6944097Z Autotune Choices Stats:
2025-12-04T09:45:16.6944839Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.6945004Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6945120Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6945279Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6945896Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6946495Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6947113Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6947717Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6948338Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6948960Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6949584Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6950192Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6950832Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6951455Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6951583Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.6951624Z Autotune Choices Stats:
2025-12-04T09:45:16.6952377Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.6952612Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6952778Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6953067Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6953710Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6954337Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6954968Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6955610Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6956238Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6956882Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6957520Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6958161Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6958788Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6959435Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6959566Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.6959638Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6959693Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6959733Z unimplemented []
2025-12-04T09:45:16.6959794Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6959894Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6960515Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6960567Z graph_break []
2025-12-04T09:45:16.6960642Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6960682Z Autotune Choices Stats:
2025-12-04T09:45:16.6961421Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.6961561Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6961675Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6961838Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6962462Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6963074Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6963684Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6964304Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6964909Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6965527Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6966140Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6966766Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6967369Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6967971Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6968102Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.6968142Z Autotune Choices Stats:
2025-12-04T09:45:16.6968922Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.6969149Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6969315Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6969594Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6970220Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6970901Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6971528Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6972159Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6972817Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6973445Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6974083Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6974717Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6975369Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6975996Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6976125Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.6976199Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6976240Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6976278Z unimplemented []
2025-12-04T09:45:16.6976338Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6976441Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6977038Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.6977076Z graph_break []
2025-12-04T09:45:16.6977150Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6977189Z Autotune Choices Stats:
2025-12-04T09:45:16.6977933Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.6978074Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6978188Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6978348Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6978982Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6979590Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6980197Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6980837Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6981456Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6982063Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6982679Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6983289Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6983922Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6984530Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6984660Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.6984700Z Autotune Choices Stats:
2025-12-04T09:45:16.6985477Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.6985701Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6985868Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6986155Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6986791Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6987418Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6988073Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6988706Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6989332Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6989976Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6990635Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6991279Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6991916Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6992575Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6992705Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.6992779Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.6992822Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.6992861Z unimplemented []
2025-12-04T09:45:16.6992921Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.6993024Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.6993604Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.6993642Z graph_break []
2025-12-04T09:45:16.6993716Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.6993756Z Autotune Choices Stats:
2025-12-04T09:45:16.6994506Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.6994643Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.6994758Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.6994921Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.6995529Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6996153Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6996759Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6997385Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6997982Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6998594Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.6999198Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.6999817Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7000460Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7001075Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7001205Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.7001245Z Autotune Choices Stats:
2025-12-04T09:45:16.7002023Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.7002242Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7002421Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7002699Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7003333Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7003998Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7004626Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7005272Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7005905Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7006539Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7007177Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7007807Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7008450Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7009097Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7009226Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.7009299Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7009342Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7009379Z unimplemented []
2025-12-04T09:45:16.7009440Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7009542Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7010116Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7010152Z graph_break []
2025-12-04T09:45:16.7010225Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7010265Z Autotune Choices Stats:
2025-12-04T09:45:16.7011064Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.7011195Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7011309Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7011469Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7012104Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7012714Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7013348Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7013949Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7014558Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7015193Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7015805Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7016412Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7017032Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7017663Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7017791Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.7017831Z Autotune Choices Stats:
2025-12-04T09:45:16.7018601Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.7018822Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7018988Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7019271Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7019931Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7020594Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7021236Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7021890Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7022521Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7023153Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7023800Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7024448Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7025078Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7025714Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7025853Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.7025927Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7025969Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7026009Z unimplemented []
2025-12-04T09:45:16.7026069Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7026183Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7026762Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7026800Z graph_break []
2025-12-04T09:45:16.7026874Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7026914Z Autotune Choices Stats:
2025-12-04T09:45:16.7027666Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.7027795Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7027920Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7028082Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7028701Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7029319Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7029927Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7030594Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7031201Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7031810Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7032429Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7033042Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7033663Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7034266Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7034408Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.7034448Z Autotune Choices Stats:
2025-12-04T09:45:16.7035231Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.7035451Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7035617Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7035898Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7036545Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7037175Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7037809Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7038448Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7039098Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7039731Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7040358Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7041030Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7041659Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7042304Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7042432Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.7042505Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7042549Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7042586Z unimplemented []
2025-12-04T09:45:16.7042662Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7042761Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7043355Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7043393Z graph_break []
2025-12-04T09:45:16.7043467Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7043506Z Autotune Choices Stats:
2025-12-04T09:45:16.7044255Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.7044385Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7044500Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7044661Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7045300Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7045908Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7046528Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7047133Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7047761Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7048368Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7048986Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7049626Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7050236Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7050890Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7051021Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.7051061Z Autotune Choices Stats:
2025-12-04T09:45:16.7051830Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.7052086Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7052253Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7052531Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7053171Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7053814Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7054440Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7055083Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7055741Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7056397Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7057023Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7057655Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7058296Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7058924Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7059061Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.7059136Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7059177Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7059217Z unimplemented []
2025-12-04T09:45:16.7059276Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7059379Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7059965Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7060013Z graph_break []
2025-12-04T09:45:16.7060087Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7060127Z Autotune Choices Stats:
2025-12-04T09:45:16.7060919Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.7061048Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7061163Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7061323Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7061933Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7062558Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7063167Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7063785Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7064388Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7065016Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7065628Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7066239Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7066862Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7067467Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7067614Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.7067654Z Autotune Choices Stats:
2025-12-04T09:45:16.7068424Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.7068656Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7068828Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7069114Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7069752Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7070382Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7071065Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7071695Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7072343Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7072991Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7073646Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7074274Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7074912Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7075555Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7075685Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.7075757Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7075800Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7075840Z unimplemented []
2025-12-04T09:45:16.7075911Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7076011Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7076593Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7076631Z graph_break []
2025-12-04T09:45:16.7076705Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7076746Z Autotune Choices Stats:
2025-12-04T09:45:16.7077505Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.7077645Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7077760Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7077920Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7078552Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7079159Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7079784Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7080392Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7081045Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7081644Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7082282Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7082894Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7083504Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7084127Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7084256Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.7084296Z Autotune Choices Stats:
2025-12-04T09:45:16.7085073Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.7085305Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7085471Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7085760Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7086409Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7087038Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7087664Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7088302Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7088932Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7089572Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7090203Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7090890Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7091524Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7092151Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7092278Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.7092368Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7092410Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7092449Z unimplemented []
2025-12-04T09:45:16.7092509Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7092610Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7093185Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7093237Z graph_break []
2025-12-04T09:45:16.7093314Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7093354Z Autotune Choices Stats:
2025-12-04T09:45:16.7094103Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.7094248Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7094364Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7094536Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7095147Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7095758Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7096382Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7096993Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7097599Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7098212Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7098832Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7099467Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7100077Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7100732Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7100862Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.7100916Z Autotune Choices Stats:
2025-12-04T09:45:16.7101682Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.7101911Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7102080Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7102361Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7102997Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7103649Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7104279Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7104907Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7105549Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7106180Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7106824Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7107458Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7108102Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7108727Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7108856Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.7108930Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7108972Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7109011Z unimplemented []
2025-12-04T09:45:16.7109073Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7109172Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7109761Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7109799Z graph_break []
2025-12-04T09:45:16.7109872Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7109913Z Autotune Choices Stats:
2025-12-04T09:45:16.7110703Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.7110844Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7110958Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7111118Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7111762Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7112368Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7112979Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7113586Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7114205Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7114810Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7115433Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7116068Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7116669Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7117275Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7117404Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.7117443Z Autotune Choices Stats:
2025-12-04T09:45:16.7118225Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.7118446Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7118610Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7118900Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7119535Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7120175Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7120853Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7121483Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7122115Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7122770Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7123397Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7124031Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7124685Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7125308Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7125437Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.7125513Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7125554Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7125594Z unimplemented []
2025-12-04T09:45:16.7125654Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7125754Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7126335Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7126374Z graph_break []
2025-12-04T09:45:16.7126458Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7126499Z Autotune Choices Stats:
2025-12-04T09:45:16.7127252Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:16.7127389Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7127505Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7127666Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7128289Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7128924Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7129535Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7130143Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7130807Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7131419Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7132051Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7132656Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7133294Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7133900Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7134031Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:16.7134073Z Autotune Choices Stats:
2025-12-04T09:45:16.7134835Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:16.7135053Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7135230Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7135510Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7136151Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7136791Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7137441Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7138069Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7138705Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7139347Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7139977Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7140647Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7141289Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7141943Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7142072Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:16.7142145Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7142190Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7142228Z unimplemented []
2025-12-04T09:45:16.7142288Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7142387Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7142967Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7143004Z graph_break []
2025-12-04T09:45:16.7143078Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7143118Z Autotune Choices Stats:
2025-12-04T09:45:16.7143872Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.7144002Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7144116Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7144289Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7144907Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7145512Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7146139Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7146745Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7147372Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7148002Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7148607Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7149226Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7149836Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7150500Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7150630Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:16.7150670Z Autotune Choices Stats:
2025-12-04T09:45:16.7151429Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.7151652Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7151816Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7152113Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7152744Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7153383Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7154012Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7154663Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7155293Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7155921Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7156561Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7157189Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7157826Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7158459Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7158597Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:16.7158672Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7158713Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7158752Z unimplemented []
2025-12-04T09:45:16.7158823Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7158924Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7159502Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7159542Z graph_break []
2025-12-04T09:45:16.7159617Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7159657Z Autotune Choices Stats:
2025-12-04T09:45:16.7160398Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.7160554Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7160683Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7160847Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7161465Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7162088Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7162698Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7163335Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7163938Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7164544Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7165184Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7165783Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7166405Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7167011Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7167153Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:16.7167194Z Autotune Choices Stats:
2025-12-04T09:45:16.7167959Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.7168177Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7168345Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7168620Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7169268Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7169900Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7170563Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7171189Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7171852Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7172482Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7173113Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7173753Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7174387Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7175030Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7175159Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:16.7175232Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7175289Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7175327Z unimplemented []
2025-12-04T09:45:16.7175388Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7175487Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7176073Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7176110Z graph_break []
2025-12-04T09:45:16.7176185Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7176225Z Autotune Choices Stats:
2025-12-04T09:45:16.7176970Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:16.7177099Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7177212Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7177377Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7178013Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7178622Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7179239Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7179846Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7180512Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7181121Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7181734Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7182356Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7182964Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7183581Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7183711Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:16.7183751Z Autotune Choices Stats:
2025-12-04T09:45:16.7184540Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.7184764Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7184930Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7185211Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7185852Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7186491Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7187117Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7187754Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7188399Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7189048Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7189676Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7190309Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7190977Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7191603Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7191751Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:16.7191827Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7191868Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7191908Z unimplemented []
2025-12-04T09:45:16.7191968Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7192068Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7192647Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7192700Z graph_break []
2025-12-04T09:45:16.7192773Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7192814Z Autotune Choices Stats:
2025-12-04T09:45:16.7193573Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:16.7193702Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7193820Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7193983Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7194601Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7195214Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7195823Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7196449Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7197069Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7197702Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7198314Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7198929Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7199548Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7200155Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7200295Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:16.7200336Z Autotune Choices Stats:
2025-12-04T09:45:16.7201132Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.7201366Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7201546Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7201824Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7202453Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7203082Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7203721Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7204347Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7204996Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7205631Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7206278Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7206903Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7207539Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7208185Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7208314Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:16.7208388Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7208442Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7208479Z unimplemented []
2025-12-04T09:45:16.7208540Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7208640Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7209218Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7209255Z graph_break []
2025-12-04T09:45:16.7209328Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7209381Z Autotune Choices Stats:
2025-12-04T09:45:16.7210138Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:16.7210267Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7210381Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7210574Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7211191Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7211819Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7212450Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7213056Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7213675Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7214295Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7214917Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7215526Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7216131Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7216752Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7216884Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:16.7216924Z Autotune Choices Stats:
2025-12-04T09:45:16.7217693Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:16.7217913Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7218078Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7218369Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7219020Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7219648Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7220271Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7220943Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7221575Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7222216Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7222855Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7223514Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7224145Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7224769Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7224910Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:16.7224986Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7225029Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7225067Z unimplemented []
2025-12-04T09:45:16.7225128Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7225230Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7225812Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7225861Z graph_break []
2025-12-04T09:45:16.7225933Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7225974Z Autotune Choices Stats:
2025-12-04T09:45:16.7226718Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.7226856Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7226972Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7227147Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7227769Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7228395Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7229011Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7229621Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7230245Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7230892Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7231533Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7232143Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7232759Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7233370Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7233512Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:16.7233555Z Autotune Choices Stats:
2025-12-04T09:45:16.7234312Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.7234545Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7234712Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7234992Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7235635Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7236274Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7236907Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7237535Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7238173Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7238805Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7239442Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7240092Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7240764Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7241395Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7241527Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:16.7241601Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7241645Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7241682Z unimplemented []
2025-12-04T09:45:16.7241743Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7241841Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7242427Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7242463Z graph_break []
2025-12-04T09:45:16.7242537Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7242590Z Autotune Choices Stats:
2025-12-04T09:45:16.7243348Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.7243476Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7243589Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7243771Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7244400Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7245010Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7245626Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7246245Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7246955Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7247574Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7248188Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7248823Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7249430Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7250040Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7250172Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:16.7250214Z Autotune Choices Stats:
2025-12-04T09:45:16.7251037Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.7251259Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7251423Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7251718Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7252355Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7253007Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7253632Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7254260Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7254910Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7255539Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7256179Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7256815Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7257468Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7258093Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7258225Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:16.7258300Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7260782Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7260824Z unimplemented []
2025-12-04T09:45:16.7260892Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7260996Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7261604Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7261647Z graph_break []
2025-12-04T09:45:16.7261722Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7261767Z Autotune Choices Stats:
2025-12-04T09:45:16.7262521Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1708", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.7262678Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7262796Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7262959Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7263574Z   triton_flex_attention_1708 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7264216Z   triton_flex_attention_1706 0.0107 ms 93.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7264824Z   triton_flex_attention_1709 0.0110 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7265429Z   triton_flex_attention_1704 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7266049Z   triton_flex_attention_1707 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7266657Z   triton_flex_attention_1705 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7267281Z   triton_flex_attention_1724 0.0146 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7267884Z   triton_flex_attention_1716 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7268520Z   triton_flex_attention_1722 0.0160 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7269127Z   triton_flex_attention_1702 0.0166 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7269260Z SingleProcess AUTOTUNE benchmarking takes 0.1988 seconds and 0.5275 seconds precompiling for 24 choices
2025-12-04T09:45:16.7269302Z Autotune Choices Stats:
2025-12-04T09:45:16.7270068Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1743", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:16.7270301Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7270575Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7270863Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7271510Z   triton_flex_attention_backward_1743 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7272140Z   triton_flex_attention_backward_1737 0.0208 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7272795Z   triton_flex_attention_backward_1734 0.0213 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7273423Z   triton_flex_attention_backward_1735 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7274059Z   triton_flex_attention_backward_1745 0.0232 ms 76.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7274702Z   triton_flex_attention_backward_1744 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7275332Z   triton_flex_attention_backward_1742 0.0249 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7275969Z   triton_flex_attention_backward_1747 0.0252 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7276598Z   triton_flex_attention_backward_1738 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7277252Z   triton_flex_attention_backward_1729 0.0264 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7277382Z SingleProcess AUTOTUNE benchmarking takes 0.2428 seconds and 0.7372 seconds precompiling for 22 choices
2025-12-04T09:45:16.7277478Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.7277526Z Traceback (most recent call last):
2025-12-04T09:45:16.7277683Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.7277724Z     self.assertTrue(
2025-12-04T09:45:16.7277832Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.7277880Z     raise self.failureException(msg)
2025-12-04T09:45:16.7278011Z AssertionError: False is not true : Log file /tmp/tmp0a8luhxf/flex_attention_configs.json was not created
2025-12-04T09:45:16.7278015Z 
2025-12-04T09:45:16.7278093Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.7278258Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.7278260Z 
2025-12-04T09:45:16.7278350Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.7278427Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7278472Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7278511Z unimplemented []
2025-12-04T09:45:16.7278574Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7279181Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.7279282Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7279335Z graph_break []
2025-12-04T09:45:16.7279410Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7279917Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.7279966Z   current_size = base.storage().size()
2025-12-04T09:45:16.7280007Z Autotune Choices Stats:
2025-12-04T09:45:16.7280789Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.7280937Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7281066Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7281229Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7281846Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7282452Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7283069Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7283674Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7284292Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7284893Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7285524Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7286129Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7286753Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7287384Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7287517Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.7287559Z Autotune Choices Stats:
2025-12-04T09:45:16.7288317Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.7288550Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7288720Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7289000Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7289660Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7290287Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7290967Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7291590Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7292231Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7292860Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7293498Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7294161Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7294789Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7295429Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7295557Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.7295634Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7295676Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7295714Z unimplemented []
2025-12-04T09:45:16.7295775Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7295887Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7296467Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7296505Z graph_break []
2025-12-04T09:45:16.7296579Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7296632Z Autotune Choices Stats:
2025-12-04T09:45:16.7297385Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.7297513Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7297628Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7297802Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7298426Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7299038Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7299660Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7300293Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7300933Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7301551Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7302161Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7302791Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7303394Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7304002Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7304134Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.7304175Z Autotune Choices Stats:
2025-12-04T09:45:16.7304953Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.7305172Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7305340Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7305628Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7306250Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7306896Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7307517Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7308159Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7308796Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7309419Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7310053Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7310713Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7311378Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7312000Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7312131Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.7312205Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7312249Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7312286Z unimplemented []
2025-12-04T09:45:16.7312348Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7312451Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7313039Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7313079Z graph_break []
2025-12-04T09:45:16.7313152Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7313195Z Autotune Choices Stats:
2025-12-04T09:45:16.7313935Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.7314078Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7314195Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7314356Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7314968Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7315591Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7316196Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7316793Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7317405Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7318014Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7318632Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7319236Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7319863Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7320491Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7320626Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.7320666Z Autotune Choices Stats:
2025-12-04T09:45:16.7321422Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.7321658Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7321825Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7322103Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7322750Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7323377Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7324026Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7324655Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7325285Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7325925Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7326556Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7327201Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7327828Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7328475Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7328604Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.7328678Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7328722Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7328760Z unimplemented []
2025-12-04T09:45:16.7328821Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7328921Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7329512Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7329550Z graph_break []
2025-12-04T09:45:16.7329622Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7329664Z Autotune Choices Stats:
2025-12-04T09:45:16.7330457Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.7330584Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7330700Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7330875Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7331488Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7332096Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7332726Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7333330Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7333929Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7334543Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7335152Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7335768Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7336378Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7337001Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7337130Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.7337172Z Autotune Choices Stats:
2025-12-04T09:45:16.7337932Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.7338149Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7338315Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7338608Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7339242Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7339879Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7340561Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7341212Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7341841Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7342488Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7343130Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7343758Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7344400Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7345037Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7345178Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.7345252Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7345305Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7345343Z unimplemented []
2025-12-04T09:45:16.7345404Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7345504Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7346083Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7346122Z graph_break []
2025-12-04T09:45:16.7346195Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7346235Z Autotune Choices Stats:
2025-12-04T09:45:16.7346976Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.7347114Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7347228Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7347391Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7347999Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7348609Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7349217Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7349839Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7350463Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7351061Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7351683Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7352289Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7352910Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7353516Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7353660Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.7353700Z Autotune Choices Stats:
2025-12-04T09:45:16.7354472Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.7354693Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7354861Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7355139Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7355788Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7356412Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7357048Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7357688Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7358339Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7358965Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7359590Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7360251Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7360905Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7361544Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7361672Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.7361747Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7361802Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7361841Z unimplemented []
2025-12-04T09:45:16.7361901Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7362002Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7362594Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7362633Z graph_break []
2025-12-04T09:45:16.7362706Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7362749Z Autotune Choices Stats:
2025-12-04T09:45:16.7363496Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.7363621Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7363736Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7363898Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7364544Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7365149Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7365768Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7366371Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7366999Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7367604Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7368210Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7368830Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7369437Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7370052Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7370181Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.7370222Z Autotune Choices Stats:
2025-12-04T09:45:16.7371042Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.7371263Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7371430Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7371709Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7372349Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7372985Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7373613Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7374250Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7374878Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7375533Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7376159Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7376802Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7377431Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7378053Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7378194Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.7378267Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7378310Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7378348Z unimplemented []
2025-12-04T09:45:16.7378408Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7378510Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7379081Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7379130Z graph_break []
2025-12-04T09:45:16.7379203Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7379242Z Autotune Choices Stats:
2025-12-04T09:45:16.7379995Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.7380125Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7380239Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7380430Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7381047Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7381671Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7382282Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7382903Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7383512Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7384148Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7384761Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7385370Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7385989Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7386595Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7386738Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.7386779Z Autotune Choices Stats:
2025-12-04T09:45:16.7387549Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.7387778Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7387953Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7388232Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7388866Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7389494Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7390131Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7390787Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7391437Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7392066Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7392714Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7393344Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7393974Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7394616Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7394744Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.7394819Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7394871Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7394909Z unimplemented []
2025-12-04T09:45:16.7394969Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7395068Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7395653Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7395690Z graph_break []
2025-12-04T09:45:16.7395763Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7395814Z Autotune Choices Stats:
2025-12-04T09:45:16.7396566Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.7396693Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7396808Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7396969Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7397584Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7398191Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7398810Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7399415Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7400032Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7400684Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7401308Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7401917Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7402524Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7403141Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7403268Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.7403309Z Autotune Choices Stats:
2025-12-04T09:45:16.7404086Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.7404306Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7404475Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7404769Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7405405Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7406033Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7406659Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7407297Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7407922Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7408561Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7409184Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7409833Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7410495Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7411118Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7411260Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.7411335Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7411378Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7411416Z unimplemented []
2025-12-04T09:45:16.7411477Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7411576Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7412160Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7412212Z graph_break []
2025-12-04T09:45:16.7412286Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7412326Z Autotune Choices Stats:
2025-12-04T09:45:16.7413072Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.7413212Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7413327Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7413502Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7414110Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7414717Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7415337Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7415943Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7416550Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7417164Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7417795Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7418399Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7419003Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7419610Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7419754Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.7419794Z Autotune Choices Stats:
2025-12-04T09:45:16.7420581Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.7420816Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7420983Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7421264Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7421897Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7422565Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7423194Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7423820Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7424469Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7425097Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7425729Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7426378Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7427005Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7427641Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7427769Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.7427842Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7427886Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7427922Z unimplemented []
2025-12-04T09:45:16.7427983Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7428082Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7428678Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7428716Z graph_break []
2025-12-04T09:45:16.7428788Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7428839Z Autotune Choices Stats:
2025-12-04T09:45:16.7429584Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.7429713Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7429828Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7429997Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7430664Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7431269Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7431895Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7432517Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7433124Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7433744Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7434357Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7434986Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7435592Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7436199Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7436327Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.7436368Z Autotune Choices Stats:
2025-12-04T09:45:16.7437143Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.7437361Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7437527Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7437817Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7438452Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7439113Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7439736Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7440380Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7441051Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7441699Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7442321Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7442963Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7443618Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7444242Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7444372Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.7444446Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7444487Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7444526Z unimplemented []
2025-12-04T09:45:16.7444586Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7444688Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7445268Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7445307Z graph_break []
2025-12-04T09:45:16.7445394Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7445434Z Autotune Choices Stats:
2025-12-04T09:45:16.7446186Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.7446327Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7446441Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7446605Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7447217Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7447847Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7448451Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7449069Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7449697Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7450305Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7450961Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7451565Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7452199Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7452801Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7452934Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.7452974Z Autotune Choices Stats:
2025-12-04T09:45:16.7453741Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.7453971Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7454139Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7454421Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7455055Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7455700Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7456342Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7456968Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7457598Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7458236Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7458863Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7459507Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7460129Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7460816Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7460944Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.7461019Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7461065Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7461103Z unimplemented []
2025-12-04T09:45:16.7461164Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7461265Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7461841Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7461878Z graph_break []
2025-12-04T09:45:16.7461950Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7461992Z Autotune Choices Stats:
2025-12-04T09:45:16.7462755Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.7462884Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7462998Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7463177Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7463796Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7464401Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7465034Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7465644Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7466250Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7466865Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7467478Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7468100Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7468706Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7469331Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7469461Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.7469501Z Autotune Choices Stats:
2025-12-04T09:45:16.7470264Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.7470523Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7470691Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7470990Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7471625Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7472264Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7472907Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7473560Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7474187Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7474824Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7475477Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7476108Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7476752Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7477382Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7477522Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.7477597Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7477639Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7477688Z unimplemented []
2025-12-04T09:45:16.7477749Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7477849Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7478427Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7478467Z graph_break []
2025-12-04T09:45:16.7478541Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7478581Z Autotune Choices Stats:
2025-12-04T09:45:16.7479329Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.7479458Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7479583Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7479745Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7480361Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7481052Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7481675Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7482325Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7482935Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7483542Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7484180Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7484779Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7485401Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7486009Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7486153Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.7486193Z Autotune Choices Stats:
2025-12-04T09:45:16.7486968Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.7487187Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7487355Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7487638Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7488301Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7488927Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7489560Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7490189Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7490883Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7491514Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7492157Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7492803Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7493432Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7494071Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7494203Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.7494279Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7494336Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7494373Z unimplemented []
2025-12-04T09:45:16.7494433Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7494533Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7495130Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7495168Z graph_break []
2025-12-04T09:45:16.7495240Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7495282Z Autotune Choices Stats:
2025-12-04T09:45:16.7496025Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.7496153Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7496268Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7496432Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7497063Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7497667Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7498285Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7498905Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7499533Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7500136Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7500795Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7501420Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7502029Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7503545Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7503676Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.7503717Z Autotune Choices Stats:
2025-12-04T09:45:16.7504493Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.7504743Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7504910Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7505195Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7505825Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7506457Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7507093Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7507730Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7508431Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7509096Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7509723Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7510372Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7511055Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7511679Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7511840Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.7511916Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7511962Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7512002Z unimplemented []
2025-12-04T09:45:16.7512064Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7512338Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7512916Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7512973Z graph_break []
2025-12-04T09:45:16.7513046Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7513087Z Autotune Choices Stats:
2025-12-04T09:45:16.7513863Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.7513991Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7514108Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7514273Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7514894Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7515500Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7516108Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7516744Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7517349Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7517985Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7518596Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7519207Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7519818Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7520469Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7520628Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.7520670Z Autotune Choices Stats:
2025-12-04T09:45:16.7521441Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.7521674Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7521856Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7522136Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7522767Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7523397Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7524023Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7524649Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7525316Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7525946Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7526601Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7527232Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7527882Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7528524Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7528657Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.7528731Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7528791Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7528828Z unimplemented []
2025-12-04T09:45:16.7528891Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7528992Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7529587Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7529624Z graph_break []
2025-12-04T09:45:16.7529700Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7529740Z Autotune Choices Stats:
2025-12-04T09:45:16.7530568Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.7530698Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7530813Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7530977Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7531594Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7532204Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7532889Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7533549Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7534210Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7534830Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7535463Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7536073Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7536694Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7537299Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7537429Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.7537469Z Autotune Choices Stats:
2025-12-04T09:45:16.7538244Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.7538488Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7538655Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7538943Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7539583Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7540210Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7540871Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7541501Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7542131Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7542798Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7543421Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7544083Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7544715Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7545343Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7545474Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.7545550Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7545592Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7545630Z unimplemented []
2025-12-04T09:45:16.7545692Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7545794Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7546369Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7546417Z graph_break []
2025-12-04T09:45:16.7546490Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7546530Z Autotune Choices Stats:
2025-12-04T09:45:16.7547289Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.7547429Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7547545Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7547716Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7548331Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7548938Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7549546Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7550154Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7550794Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7551450Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7552081Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7552707Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7553312Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7553917Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7554049Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.7554092Z Autotune Choices Stats:
2025-12-04T09:45:16.7554856Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.7555102Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7555269Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7555560Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7556191Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7556840Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7557469Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7558091Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7558717Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7559346Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7559998Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7560664Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7561306Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7561934Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7562063Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.7562137Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7562181Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7562219Z unimplemented []
2025-12-04T09:45:16.7562281Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7562381Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7562953Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7562989Z graph_break []
2025-12-04T09:45:16.7563062Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7563102Z Autotune Choices Stats:
2025-12-04T09:45:16.7563864Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.7564007Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7564120Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7564295Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7564929Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7565534Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7566142Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7566744Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7567340Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7567939Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7568568Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7569195Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7569798Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7570447Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7570578Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.7570618Z Autotune Choices Stats:
2025-12-04T09:45:16.7571386Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.7571610Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7571775Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7572067Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7572716Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7573359Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7573995Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7574626Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7575258Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7575903Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7576536Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7577191Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7577832Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7578458Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7578588Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.7578661Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7578703Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7578739Z unimplemented []
2025-12-04T09:45:16.7578801Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7578901Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7579484Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7579522Z graph_break []
2025-12-04T09:45:16.7579595Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7579635Z Autotune Choices Stats:
2025-12-04T09:45:16.7580379Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.7580545Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7580659Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7580821Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7581454Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7582087Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7582690Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7583300Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7583909Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7584518Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7585140Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7585755Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7586381Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7586989Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7587118Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.7587159Z Autotune Choices Stats:
2025-12-04T09:45:16.7587925Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.7588145Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7588312Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7588591Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7589225Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7589876Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7590557Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7591183Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7591813Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7592461Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7593084Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7593711Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7594358Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7595006Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7595136Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.7595210Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7595251Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7595292Z unimplemented []
2025-12-04T09:45:16.7595352Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7595455Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7596031Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7596068Z graph_break []
2025-12-04T09:45:16.7596142Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7596181Z Autotune Choices Stats:
2025-12-04T09:45:16.7596928Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.7597056Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7597170Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7597343Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7597948Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7598567Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7599278Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7599880Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7600504Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7601105Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7601714Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7602332Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7602947Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7603575Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7603706Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.7603745Z Autotune Choices Stats:
2025-12-04T09:45:16.7604516Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.7604734Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7604900Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7605181Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7605815Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7606441Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7607086Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7607723Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7608350Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7608974Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7609611Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7610261Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7610922Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7611562Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7611703Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.7611776Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7611819Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7611856Z unimplemented []
2025-12-04T09:45:16.7611932Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7612032Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7612605Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7612643Z graph_break []
2025-12-04T09:45:16.7612716Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7612757Z Autotune Choices Stats:
2025-12-04T09:45:16.7613503Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.7613632Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7613745Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7613905Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7614513Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7615126Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7615741Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7616375Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7616975Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7617598Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7618208Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7618814Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7619429Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7620041Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7620180Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.7620221Z Autotune Choices Stats:
2025-12-04T09:45:16.7621023Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.7621243Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7621418Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7621708Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7622341Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7622967Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7623603Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7624242Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7624891Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7625518Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7626140Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7626790Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7627438Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7628072Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7628212Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.7628287Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7628328Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7628378Z unimplemented []
2025-12-04T09:45:16.7628438Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7628538Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7629122Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7629160Z graph_break []
2025-12-04T09:45:16.7629235Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7629274Z Autotune Choices Stats:
2025-12-04T09:45:16.7630013Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.7630140Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7630254Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7630441Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7631054Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7631659Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7632300Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7632901Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7633531Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7634134Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7634739Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7635361Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7635969Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7636582Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7636724Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.7636764Z Autotune Choices Stats:
2025-12-04T09:45:16.7637529Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.7637759Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7637929Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7638208Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7638846Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7639465Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7640095Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7640766Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7641410Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7642069Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7642692Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7643321Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7643952Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7644577Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7644720Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.7644793Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7644836Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7644873Z unimplemented []
2025-12-04T09:45:16.7644935Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7645037Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7645629Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7645677Z graph_break []
2025-12-04T09:45:16.7645749Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7645789Z Autotune Choices Stats:
2025-12-04T09:45:16.7646548Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.7646676Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7646789Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7646953Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7647576Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7648182Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7648788Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7649418Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7650025Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7650677Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7651290Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7651900Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7652510Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7653117Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7653261Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.7653303Z Autotune Choices Stats:
2025-12-04T09:45:16.7654081Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.7654315Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7654493Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7654771Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7655412Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7656044Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7656671Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7657289Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7657938Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7658564Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7659210Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7659839Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7660482Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7661113Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7661241Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.7661316Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7661358Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7661411Z unimplemented []
2025-12-04T09:45:16.7661470Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7661571Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7662154Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7662192Z graph_break []
2025-12-04T09:45:16.7662265Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7662304Z Autotune Choices Stats:
2025-12-04T09:45:16.7663070Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.7663198Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7663312Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7663475Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7664091Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7664692Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7665302Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7665897Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7666524Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7667131Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7667769Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7668377Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7668987Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7669600Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7669727Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.7669768Z Autotune Choices Stats:
2025-12-04T09:45:16.7670560Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.7670806Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7670975Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7671267Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7671917Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7672543Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7673171Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7673802Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7674449Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7675102Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7675730Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7676382Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7677010Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7677636Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7677770Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.7677843Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7677889Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7677928Z unimplemented []
2025-12-04T09:45:16.7677994Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7678094Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7678666Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7678714Z graph_break []
2025-12-04T09:45:16.7678788Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7678828Z Autotune Choices Stats:
2025-12-04T09:45:16.7679587Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.7679735Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7679848Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7680021Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7680664Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7681271Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7681874Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7682483Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7683085Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7683710Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7684346Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7684956Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7685565Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7686170Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7686300Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.7686340Z Autotune Choices Stats:
2025-12-04T09:45:16.7687105Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.7687333Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7687498Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7687788Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7688422Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7689066Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7689696Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7690328Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7690980Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7691611Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7692261Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7692920Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7693550Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7694177Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7694305Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.7694379Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7694422Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7694459Z unimplemented []
2025-12-04T09:45:16.7694519Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7694621Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7695194Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7695232Z graph_break []
2025-12-04T09:45:16.7695304Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7695346Z Autotune Choices Stats:
2025-12-04T09:45:16.7696103Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.7696248Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7696362Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7696534Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7697160Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7697770Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7698380Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7698985Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7699595Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7700206Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7700891Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7701522Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7702122Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7702731Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7702858Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.7702900Z Autotune Choices Stats:
2025-12-04T09:45:16.7703664Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.7703884Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7704051Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7704340Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7704988Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7705638Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7706263Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7706904Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7707531Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7708155Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7708779Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7709431Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7710085Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7710741Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7710872Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.7710945Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7710987Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7711025Z unimplemented []
2025-12-04T09:45:16.7711086Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7711186Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7711763Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7711800Z graph_break []
2025-12-04T09:45:16.7711874Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7711913Z Autotune Choices Stats:
2025-12-04T09:45:16.7712665Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.7712815Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7712927Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7713092Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7713721Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7714350Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7714954Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7715559Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7716164Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7716766Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7717387Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7718005Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7718632Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7719236Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7719366Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.7719407Z Autotune Choices Stats:
2025-12-04T09:45:16.7720163Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.7720381Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7720575Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7720857Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7721492Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7722145Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7722798Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7723509Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7724135Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7724765Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7725395Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7726031Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7726671Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7727322Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7727450Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.7727527Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7727570Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7727609Z unimplemented []
2025-12-04T09:45:16.7727669Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7727771Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7728349Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7728387Z graph_break []
2025-12-04T09:45:16.7728460Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7728503Z Autotune Choices Stats:
2025-12-04T09:45:16.7729251Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.7729378Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7729492Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7729663Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7730285Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7730933Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7731559Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7732162Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7732769Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7733374Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7733989Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7734608Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7735226Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7735857Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7735986Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.7736028Z Autotune Choices Stats:
2025-12-04T09:45:16.7736791Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.7737008Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7737177Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7737457Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7738099Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7738737Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7739381Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7740025Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7740686Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7741319Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7741964Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7742593Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7743236Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7743878Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7744025Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.7744099Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7744155Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7744192Z unimplemented []
2025-12-04T09:45:16.7744255Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7744357Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7744941Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7744978Z graph_break []
2025-12-04T09:45:16.7745052Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7745092Z Autotune Choices Stats:
2025-12-04T09:45:16.7745847Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.7745975Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7746088Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7746253Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7746875Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7747505Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7748113Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7748732Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7749340Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7749950Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7750582Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7751181Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7751805Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7752428Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7752571Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.7752612Z Autotune Choices Stats:
2025-12-04T09:45:16.7753387Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.7753609Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7753776Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7754057Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7754685Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7755303Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7755933Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7756572Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7757228Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7757847Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7758483Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7759131Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7759777Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7760459Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7760587Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.7760661Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7760717Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7760756Z unimplemented []
2025-12-04T09:45:16.7760816Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7760917Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7761508Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7761546Z graph_break []
2025-12-04T09:45:16.7761619Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7761662Z Autotune Choices Stats:
2025-12-04T09:45:16.7762418Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:16.7762546Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7762661Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7762822Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7763439Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7766187Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7766838Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7767445Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7768079Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7768680Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7769291Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7769902Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7770550Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7771178Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7771309Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:16.7771365Z Autotune Choices Stats:
2025-12-04T09:45:16.7772155Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:16.7772376Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7772547Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7772831Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7773470Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7774115Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7774750Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7775401Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7776029Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7776683Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7777309Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7777938Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7778567Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7779199Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7779346Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:16.7779422Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7779466Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7779505Z unimplemented []
2025-12-04T09:45:16.7779567Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7779680Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7780272Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7780321Z graph_break []
2025-12-04T09:45:16.7780397Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7780493Z Autotune Choices Stats:
2025-12-04T09:45:16.7781241Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.7781372Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7781487Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7781651Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7782273Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7782884Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7783508Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7784152Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7784771Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7785390Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7786002Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7786616Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7787243Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7787851Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7787992Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:16.7788032Z Autotune Choices Stats:
2025-12-04T09:45:16.7788800Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.7789031Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7789206Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7789486Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7790122Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7790795Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7791422Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7792051Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7792713Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7793341Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7793994Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7794629Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7795259Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7795892Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7796021Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:16.7796097Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7796150Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7796188Z unimplemented []
2025-12-04T09:45:16.7796249Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7796351Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7796941Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7796979Z graph_break []
2025-12-04T09:45:16.7797052Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7797105Z Autotune Choices Stats:
2025-12-04T09:45:16.7797858Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.7797985Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7798100Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7798263Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7798882Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7799488Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7800114Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7800770Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7801397Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7802026Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7802635Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7803250Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7803864Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7804475Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7804603Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:16.7804660Z Autotune Choices Stats:
2025-12-04T09:45:16.7805419Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.7805649Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7805827Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7806109Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7806761Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7807393Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7808023Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7808650Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7809295Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7809948Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7810643Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7811270Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7811908Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7812536Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7812666Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:16.7812742Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7812784Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7812822Z unimplemented []
2025-12-04T09:45:16.7812882Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7812982Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7813559Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7813610Z graph_break []
2025-12-04T09:45:16.7813684Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7813724Z Autotune Choices Stats:
2025-12-04T09:45:16.7814491Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:16.7814632Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7814757Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7814920Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7815537Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7816149Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7816759Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7817368Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7817984Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7818594Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7819227Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7819833Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7820473Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7821076Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7821207Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:16.7821246Z Autotune Choices Stats:
2025-12-04T09:45:16.7822012Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.7822244Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7822424Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7822700Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7823362Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7823990Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7824617Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7825259Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7825888Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7826517Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7827172Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7827822Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7828451Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7829084Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7829211Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:16.7829285Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7829328Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7829364Z unimplemented []
2025-12-04T09:45:16.7829425Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7829525Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7830113Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7830151Z graph_break []
2025-12-04T09:45:16.7830224Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7830279Z Autotune Choices Stats:
2025-12-04T09:45:16.7831068Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:16.7831197Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7831312Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7831490Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7832121Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7832728Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7833350Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7833960Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7834571Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7835192Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7835816Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7836447Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7837057Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7837663Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7837791Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:16.7837834Z Autotune Choices Stats:
2025-12-04T09:45:16.7838602Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.7838822Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7838998Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7839277Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7839920Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7840604Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7841229Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7841849Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7842477Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7843109Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7843751Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7844412Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7845067Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7845693Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7845823Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:16.7845898Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7845941Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7845979Z unimplemented []
2025-12-04T09:45:16.7846040Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7846142Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7846720Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7846758Z graph_break []
2025-12-04T09:45:16.7846833Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7846873Z Autotune Choices Stats:
2025-12-04T09:45:16.7847621Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:16.7847761Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7847875Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7848048Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7848665Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7849290Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7849901Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7850558Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7851162Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7851771Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7852398Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7853020Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7853653Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7854259Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7854391Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:16.7854432Z Autotune Choices Stats:
2025-12-04T09:45:16.7855200Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:16.7855418Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7855585Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7855866Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7856514Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7857153Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7857799Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7858425Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7859061Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7859695Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7860328Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7861001Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7861650Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7862303Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7862432Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:16.7862508Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7862551Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7862588Z unimplemented []
2025-12-04T09:45:16.7862648Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7862748Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7863323Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7863361Z graph_break []
2025-12-04T09:45:16.7863435Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7863476Z Autotune Choices Stats:
2025-12-04T09:45:16.7864226Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.7864353Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7864481Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7864644Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7865273Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7865880Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7866513Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7867124Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7867737Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7868346Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7868956Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7869585Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7870197Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7870849Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7870979Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:16.7871022Z Autotune Choices Stats:
2025-12-04T09:45:16.7871787Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.7872005Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7872175Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7872450Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7873092Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7873734Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7874373Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7875027Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7875656Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7876286Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7876930Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7877562Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7878210Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7878833Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7878971Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:16.7879055Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7879097Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7879137Z unimplemented []
2025-12-04T09:45:16.7879198Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7879299Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7879878Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7879916Z graph_break []
2025-12-04T09:45:16.7879991Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7880031Z Autotune Choices Stats:
2025-12-04T09:45:16.7880801Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.7880930Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7881047Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7881211Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7881833Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7882467Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7883086Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7883704Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7884312Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7884921Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7885538Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7886141Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7886770Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7887368Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7887507Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:16.7887557Z Autotune Choices Stats:
2025-12-04T09:45:16.7888313Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.7888531Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7888699Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7888976Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7889617Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7890245Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7890937Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7891562Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7892220Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7892850Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7893478Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7894103Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7894737Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7895385Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7895514Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:16.7895599Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7895643Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7895679Z unimplemented []
2025-12-04T09:45:16.7895741Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7895840Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7896429Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7896469Z graph_break []
2025-12-04T09:45:16.7896542Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7896583Z Autotune Choices Stats:
2025-12-04T09:45:16.7897329Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1708", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.7897456Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7897571Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7897732Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7898355Z   triton_flex_attention_1708 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7898962Z   triton_flex_attention_1706 0.0107 ms 93.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7899591Z   triton_flex_attention_1709 0.0110 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7900213Z   triton_flex_attention_1704 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7900843Z   triton_flex_attention_1707 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7901449Z   triton_flex_attention_1705 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7902062Z   triton_flex_attention_1724 0.0146 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7902679Z   triton_flex_attention_1716 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7903288Z   triton_flex_attention_1722 0.0160 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7903920Z   triton_flex_attention_1702 0.0166 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7904062Z SingleProcess AUTOTUNE benchmarking takes 0.1988 seconds and 0.5275 seconds precompiling for 24 choices
2025-12-04T09:45:16.7904104Z Autotune Choices Stats:
2025-12-04T09:45:16.7904882Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1743", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:16.7905099Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7905267Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7905543Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7906179Z   triton_flex_attention_backward_1743 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7906814Z   triton_flex_attention_backward_1737 0.0208 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7907441Z   triton_flex_attention_backward_1734 0.0213 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7908091Z   triton_flex_attention_backward_1735 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7908722Z   triton_flex_attention_backward_1745 0.0232 ms 76.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7909374Z   triton_flex_attention_backward_1744 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7910004Z   triton_flex_attention_backward_1742 0.0249 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7910672Z   triton_flex_attention_backward_1747 0.0252 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7911306Z   triton_flex_attention_backward_1738 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7911933Z   triton_flex_attention_backward_1729 0.0264 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7912076Z SingleProcess AUTOTUNE benchmarking takes 0.2428 seconds and 0.7372 seconds precompiling for 22 choices
2025-12-04T09:45:16.7912153Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7912195Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7912245Z unimplemented []
2025-12-04T09:45:16.7912306Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7912407Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7912997Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7913036Z graph_break []
2025-12-04T09:45:16.7913123Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7913162Z Autotune Choices Stats:
2025-12-04T09:45:16.7913907Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1754", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.7914034Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7914150Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7914312Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7914935Z   triton_flex_attention_1754 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7915546Z   triton_flex_attention_1752 0.0110 ms 90.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7916164Z   triton_flex_attention_1755 0.0114 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7916781Z   triton_flex_attention_1753 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7917404Z   triton_flex_attention_1750 0.0125 ms 79.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7918011Z   triton_flex_attention_1751 0.0143 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7918626Z   triton_flex_attention_1770 0.0149 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7919237Z   triton_flex_attention_1762 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7919851Z   triton_flex_attention_1768 0.0163 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7920491Z   triton_flex_attention_1748 0.0170 ms 58.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7920635Z SingleProcess AUTOTUNE benchmarking takes 0.2060 seconds and 0.4503 seconds precompiling for 24 choices
2025-12-04T09:45:16.7920674Z Autotune Choices Stats:
2025-12-04T09:45:16.7921453Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1789", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.7921683Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7921868Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7922148Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7922784Z   triton_flex_attention_backward_1789 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7923411Z   triton_flex_attention_backward_1783 0.0209 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7924043Z   triton_flex_attention_backward_1780 0.0216 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7924663Z   triton_flex_attention_backward_1781 0.0217 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7925315Z   triton_flex_attention_backward_1791 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7925959Z   triton_flex_attention_backward_1790 0.0235 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7926584Z   triton_flex_attention_backward_1788 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7927221Z   triton_flex_attention_backward_1793 0.0255 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7927839Z   triton_flex_attention_backward_1775 0.0264 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7928478Z   triton_flex_attention_backward_1784 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7928607Z SingleProcess AUTOTUNE benchmarking takes 0.2498 seconds and 0.6949 seconds precompiling for 22 choices
2025-12-04T09:45:16.7928711Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.7928759Z Traceback (most recent call last):
2025-12-04T09:45:16.7928916Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.7928957Z     self.assertTrue(
2025-12-04T09:45:16.7929066Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.7929117Z     raise self.failureException(msg)
2025-12-04T09:45:16.7929242Z AssertionError: False is not true : Log file /tmp/tmp7_fehk8b/flex_attention_configs.json was not created
2025-12-04T09:45:16.7929257Z 
2025-12-04T09:45:16.7929334Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.7929500Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.7929514Z 
2025-12-04T09:45:16.7929604Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.7929680Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7929722Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7929759Z unimplemented []
2025-12-04T09:45:16.7929821Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7930457Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.7930559Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7930597Z graph_break []
2025-12-04T09:45:16.7930671Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7931168Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.7931217Z   current_size = base.storage().size()
2025-12-04T09:45:16.7931257Z Autotune Choices Stats:
2025-12-04T09:45:16.7932004Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.7932134Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7932251Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7932416Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7933033Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7933662Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7934290Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7934892Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7935491Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7936096Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7936708Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7937312Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7937930Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7938535Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7938691Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.7938733Z Autotune Choices Stats:
2025-12-04T09:45:16.7939497Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.7939720Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7939887Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7940164Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7940827Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7941454Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7942101Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7942721Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7943367Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7943990Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7944616Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7945250Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7945881Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7946523Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7946663Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.7946737Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7946781Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7946818Z unimplemented []
2025-12-04T09:45:16.7946879Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7946979Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7947570Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7947608Z graph_break []
2025-12-04T09:45:16.7947683Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7947722Z Autotune Choices Stats:
2025-12-04T09:45:16.7948462Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.7948591Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7948706Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7948869Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7949483Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7950080Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7950740Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7951365Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7951965Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7952564Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7953171Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7953781Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7954397Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7955035Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7955176Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.7955216Z Autotune Choices Stats:
2025-12-04T09:45:16.7955997Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.7956214Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7956382Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7956662Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7957315Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7957945Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7958571Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7959216Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7959847Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7960534Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7961158Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7961808Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7962436Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7963074Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7963218Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.7963295Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7963337Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7963386Z unimplemented []
2025-12-04T09:45:16.7963446Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7963547Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7964139Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7964180Z graph_break []
2025-12-04T09:45:16.7964263Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7964305Z Autotune Choices Stats:
2025-12-04T09:45:16.7965045Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.7965174Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7965290Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7965451Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7966078Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7966683Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7967295Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7967907Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7968532Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7969140Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7969746Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7970361Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7971023Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7971628Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7971779Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.7971820Z Autotune Choices Stats:
2025-12-04T09:45:16.7972601Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.7972833Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7973013Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7973290Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7973927Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7974554Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7975183Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7975810Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7976459Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7977107Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7977739Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7978362Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7978992Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7979622Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7979752Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.7979836Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7979881Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7979918Z unimplemented []
2025-12-04T09:45:16.7979979Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7980078Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7980712Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.7980764Z graph_break []
2025-12-04T09:45:16.7980838Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7980878Z Autotune Choices Stats:
2025-12-04T09:45:16.7981638Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.7981767Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7981882Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7982041Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7982656Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7983263Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7983892Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7984508Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7985137Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7985764Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7986378Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7987003Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7987614Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7988221Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7988352Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.7988402Z Autotune Choices Stats:
2025-12-04T09:45:16.7989180Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.7989399Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7989574Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7989864Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7990526Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7991166Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7991788Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7992415Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7993046Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7993700Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7994348Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7994976Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.7995606Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7996227Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7996356Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.7996431Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.7996472Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.7996511Z unimplemented []
2025-12-04T09:45:16.7996572Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.7996672Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.7997252Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.7997301Z graph_break []
2025-12-04T09:45:16.7997375Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.7997416Z Autotune Choices Stats:
2025-12-04T09:45:16.7998185Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.7998324Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.7998453Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.7998613Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.7999235Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.7999839Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8000460Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8001067Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8001685Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8002302Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8002935Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8003542Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8004155Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8004763Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8004896Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.8004937Z Autotune Choices Stats:
2025-12-04T09:45:16.8005701Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.8005931Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8006111Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8006388Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8007046Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8007672Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8008297Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8008936Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8009576Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8010215Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8010901Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8011549Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8012172Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8012802Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8012932Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.8013009Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8013051Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8013089Z unimplemented []
2025-12-04T09:45:16.8013148Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8013249Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8013831Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8013881Z graph_break []
2025-12-04T09:45:16.8013955Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8013997Z Autotune Choices Stats:
2025-12-04T09:45:16.8014752Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.8014879Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8015004Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8015166Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8015794Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8016399Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8017006Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8017612Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8018223Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8018837Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8019466Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8020094Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8020731Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8021333Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8021462Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.8021503Z Autotune Choices Stats:
2025-12-04T09:45:16.8022267Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.8022488Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8022669Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8022950Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8023598Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8024252Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8024875Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8025517Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8026144Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8026777Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8027413Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8028056Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8028702Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8029335Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8029464Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.8029538Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8029582Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8029619Z unimplemented []
2025-12-04T09:45:16.8029681Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8029780Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8030366Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8030433Z graph_break []
2025-12-04T09:45:16.8030506Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8030546Z Autotune Choices Stats:
2025-12-04T09:45:16.8031296Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.8031440Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8031554Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8031729Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8032341Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8032973Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8033577Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8034176Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8034779Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8035382Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8036006Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8036620Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8037246Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8037851Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8037982Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.8038023Z Autotune Choices Stats:
2025-12-04T09:45:16.8038775Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.8038994Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8039162Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8039441Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8040088Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8040758Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8041405Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8042032Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8042663Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8043293Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8043916Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8044560Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8045202Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8045854Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8045982Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.8046058Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8046100Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8046139Z unimplemented []
2025-12-04T09:45:16.8046198Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8046300Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8046879Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8046916Z graph_break []
2025-12-04T09:45:16.8046991Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8047031Z Autotune Choices Stats:
2025-12-04T09:45:16.8047779Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.8047907Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8048033Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8048194Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8048822Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8049426Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8050053Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8050696Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8051302Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8051904Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8052517Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8053148Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8053750Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8054381Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8054511Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.8054551Z Autotune Choices Stats:
2025-12-04T09:45:16.8055314Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.8055531Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8055697Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8055979Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8056611Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8057249Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8057895Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8058542Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8059168Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8059798Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8060442Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8061074Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8061730Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8062355Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8062498Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.8062584Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8062627Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8062664Z unimplemented []
2025-12-04T09:45:16.8062727Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8062826Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8063396Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8063435Z graph_break []
2025-12-04T09:45:16.8063508Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8063548Z Autotune Choices Stats:
2025-12-04T09:45:16.8064292Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.8064421Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8064537Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8064700Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8065316Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8065938Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8066548Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8067178Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8067782Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8068390Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8069001Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8069609Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8070235Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8070874Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8071022Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.8071078Z Autotune Choices Stats:
2025-12-04T09:45:16.8071838Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.8072057Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8072226Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8072504Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8073144Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8073768Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8074420Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8075045Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8075697Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8076328Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8076950Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8077581Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8078211Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8078861Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8078989Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.8079074Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8079114Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8079152Z unimplemented []
2025-12-04T09:45:16.8079211Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8079313Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8079903Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8079942Z graph_break []
2025-12-04T09:45:16.8080017Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8080056Z Autotune Choices Stats:
2025-12-04T09:45:16.8080847Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.8080972Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8081089Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8081250Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8081859Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8082458Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8083096Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8083710Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8084327Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8084936Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8085550Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8086159Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8086766Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8087394Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8087524Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.8087575Z Autotune Choices Stats:
2025-12-04T09:45:16.8088351Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.8088570Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8088741Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8089022Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8089656Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8090285Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8090944Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8091592Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8092224Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8092879Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8093497Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8094129Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8094759Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8095396Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8095538Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.8095612Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8095657Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8095693Z unimplemented []
2025-12-04T09:45:16.8095766Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8095866Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8096445Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8096496Z graph_break []
2025-12-04T09:45:16.8096581Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8096622Z Autotune Choices Stats:
2025-12-04T09:45:16.8097362Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.8097490Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8097604Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8097766Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8098386Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8098991Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8099589Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8100226Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8100899Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8101503Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8102112Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8102733Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8103342Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8103959Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8104120Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.8104162Z Autotune Choices Stats:
2025-12-04T09:45:16.8104924Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.8105155Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8105332Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8105609Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8106243Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8106883Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8107524Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8108142Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8108911Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8109547Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8110182Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8110834Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8111502Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8112134Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8112261Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.8112354Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8112396Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8112434Z unimplemented []
2025-12-04T09:45:16.8112494Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8112593Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8113180Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8113218Z graph_break []
2025-12-04T09:45:16.8113305Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8113344Z Autotune Choices Stats:
2025-12-04T09:45:16.8114110Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.8114236Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8114352Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8114514Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8115135Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8115740Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8116348Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8116962Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8117577Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8118199Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8118809Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8119420Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8120026Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8120653Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8120781Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.8120834Z Autotune Choices Stats:
2025-12-04T09:45:16.8121616Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.8121833Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8122014Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8122292Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8122939Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8123570Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8124195Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8124840Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8125485Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8126134Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8126777Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8127408Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8128038Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8128666Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8128796Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.8128869Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8128911Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8128947Z unimplemented []
2025-12-04T09:45:16.8129008Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8129108Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8129680Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8129727Z graph_break []
2025-12-04T09:45:16.8129803Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8129842Z Autotune Choices Stats:
2025-12-04T09:45:16.8130640Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.8130786Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8130911Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8131074Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8131686Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8132297Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8132894Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8133507Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8134123Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8134749Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8135376Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8135980Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8136603Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8137209Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8137341Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.8137380Z Autotune Choices Stats:
2025-12-04T09:45:16.8138142Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.8138370Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8138548Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8138828Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8139486Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8140110Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8140775Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8141417Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8142051Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8142681Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8143339Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8143992Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8144620Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8145248Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8145377Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.8145453Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8145494Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8145531Z unimplemented []
2025-12-04T09:45:16.8145598Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8145698Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8146280Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8146319Z graph_break []
2025-12-04T09:45:16.8146393Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8146446Z Autotune Choices Stats:
2025-12-04T09:45:16.8147196Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.8147324Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8147440Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8147612Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8148237Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8148843Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8149452Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8150057Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8150696Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8151318Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8151941Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8152577Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8153181Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8153786Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8153913Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.8153956Z Autotune Choices Stats:
2025-12-04T09:45:16.8154726Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.8154942Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8155119Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8155398Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8156045Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8156691Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8157320Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8157947Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8158576Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8159209Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8159848Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8160527Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8161184Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8161809Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8161942Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.8162017Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8162061Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8162098Z unimplemented []
2025-12-04T09:45:16.8162160Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8162258Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8162836Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8162875Z graph_break []
2025-12-04T09:45:16.8162949Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8162989Z Autotune Choices Stats:
2025-12-04T09:45:16.8163736Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.8163882Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8163996Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8164164Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8164781Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8165409Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8166024Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8166631Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8167240Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8167841Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8168464Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8169082Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8169706Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8170311Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8170471Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.8170511Z Autotune Choices Stats:
2025-12-04T09:45:16.8171274Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.8171495Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8171663Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8171941Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8172589Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8173225Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8173875Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8174501Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8175131Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8175755Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8176374Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8177016Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8177654Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8178305Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8178434Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.8178510Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8178551Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8178589Z unimplemented []
2025-12-04T09:45:16.8178649Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8178750Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8179332Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8179368Z graph_break []
2025-12-04T09:45:16.8179441Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8179484Z Autotune Choices Stats:
2025-12-04T09:45:16.8180228Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.8180355Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8180501Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8180675Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8181307Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8181912Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8182543Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8183147Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8183754Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8184373Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8184980Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8185597Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8186214Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8186835Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8186963Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.8187005Z Autotune Choices Stats:
2025-12-04T09:45:16.8187770Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.8187989Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8188155Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8188437Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8189071Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8189709Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8190343Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8191023Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8191651Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8192283Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8192910Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8193539Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8194181Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8194833Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8194980Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.8195064Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8195107Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8195144Z unimplemented []
2025-12-04T09:45:16.8195205Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8195304Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8195877Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8195914Z graph_break []
2025-12-04T09:45:16.8195988Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8196028Z Autotune Choices Stats:
2025-12-04T09:45:16.8196776Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.8196905Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8197019Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8197180Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8197798Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8198423Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8199032Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8199656Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8200261Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8200905Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8201517Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8202121Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8202744Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8203349Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8203491Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.8203530Z Autotune Choices Stats:
2025-12-04T09:45:16.8204304Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.8204523Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8204690Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8204970Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8205610Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8206232Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8206865Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8207499Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8208152Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8208778Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8209398Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8210032Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8210688Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8211344Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8211473Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.8211562Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8211604Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8211643Z unimplemented []
2025-12-04T09:45:16.8211703Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8211805Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8212396Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8212434Z graph_break []
2025-12-04T09:45:16.8212507Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8212550Z Autotune Choices Stats:
2025-12-04T09:45:16.8213296Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.8213424Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8213539Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8213701Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8214311Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8214914Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8215543Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8216148Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8216773Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8217370Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8217974Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8218583Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8219192Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8219822Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8219952Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.8220006Z Autotune Choices Stats:
2025-12-04T09:45:16.8220797Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.8221017Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8221187Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8221472Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8222110Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8222740Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8223365Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8224019Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8224648Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8225304Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8225932Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8226559Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8227182Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8227814Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8227961Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.8228036Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8228079Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8228117Z unimplemented []
2025-12-04T09:45:16.8228176Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8228287Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8228863Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8228912Z graph_break []
2025-12-04T09:45:16.8228986Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8229035Z Autotune Choices Stats:
2025-12-04T09:45:16.8229777Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.8229905Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8230020Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8230180Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8230828Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8231430Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8232040Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8232676Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8233290Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8233901Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8234511Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8235119Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8235725Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8236322Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8236463Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.8236503Z Autotune Choices Stats:
2025-12-04T09:45:16.8237279Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.8237507Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8237681Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8237960Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8238588Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8239217Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8239860Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8240538Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8241197Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8241826Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8242479Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8243111Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8243740Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8244375Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8244504Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.8244578Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8244634Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8244671Z unimplemented []
2025-12-04T09:45:16.8244731Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8244831Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8245421Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8245460Z graph_break []
2025-12-04T09:45:16.8245534Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8245587Z Autotune Choices Stats:
2025-12-04T09:45:16.8246331Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.8246459Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8246574Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8246734Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8247341Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8247954Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8248552Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8249150Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8249777Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8250443Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8251053Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8251660Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8252268Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8252890Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8253018Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.8253071Z Autotune Choices Stats:
2025-12-04T09:45:16.8253840Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.8254070Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8254237Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8254527Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8255174Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8255801Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8256427Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8257047Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8257678Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8258327Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8258967Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8259611Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8260241Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8260904Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8261032Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.8261107Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8261150Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8261188Z unimplemented []
2025-12-04T09:45:16.8261247Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8261348Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8261928Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8261980Z graph_break []
2025-12-04T09:45:16.8262053Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8262094Z Autotune Choices Stats:
2025-12-04T09:45:16.8262841Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.8262980Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8263106Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8263267Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8263883Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8264483Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8265081Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8265686Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8269825Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8270539Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8271171Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8271779Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8272386Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8272989Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8273118Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.8273157Z Autotune Choices Stats:
2025-12-04T09:45:16.8273934Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.8274162Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8274329Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8274622Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8275265Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8277700Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8278330Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8278952Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8279580Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8280203Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8280902Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8281555Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8282185Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8282815Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8282947Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.8283024Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8283071Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8283108Z unimplemented []
2025-12-04T09:45:16.8283170Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8283271Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8283851Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8283890Z graph_break []
2025-12-04T09:45:16.8283964Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8284018Z Autotune Choices Stats:
2025-12-04T09:45:16.8284760Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.8284899Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8285015Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8285187Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8285813Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8286419Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8287028Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8287647Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8288267Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8288880Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8289497Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8290121Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8290769Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8291376Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8291508Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.8291550Z Autotune Choices Stats:
2025-12-04T09:45:16.8292317Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.8292540Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8292709Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8293008Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8293655Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8294304Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8294927Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8295549Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8296181Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8296815Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8297454Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8298099Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8298748Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8299373Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8299503Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.8299579Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8299622Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8299661Z unimplemented []
2025-12-04T09:45:16.8299721Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8299824Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8300401Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8300466Z graph_break []
2025-12-04T09:45:16.8300540Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8300580Z Autotune Choices Stats:
2025-12-04T09:45:16.8301325Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.8301468Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8301585Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8301747Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8302377Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8303006Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8303611Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8304214Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8304816Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8305422Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8306040Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8306653Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8307277Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8307882Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8308012Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.8308052Z Autotune Choices Stats:
2025-12-04T09:45:16.8308822Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.8309041Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8309210Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8309498Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8310134Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8310792Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8311454Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8312079Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8312710Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8313332Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8313961Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8314604Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8315241Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8315886Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8316015Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.8316088Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8316132Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8316170Z unimplemented []
2025-12-04T09:45:16.8316231Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8316330Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8316905Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8316942Z graph_break []
2025-12-04T09:45:16.8317016Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8317057Z Autotune Choices Stats:
2025-12-04T09:45:16.8317796Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.8317927Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8318042Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8318214Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8318849Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8319455Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8320080Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8320703Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8321310Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8321915Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8322533Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8323156Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8323776Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8324413Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8324543Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.8324584Z Autotune Choices Stats:
2025-12-04T09:45:16.8325368Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.8325588Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8325754Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8326033Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8326670Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8327308Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8327943Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8328594Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8329226Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8329858Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8330523Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8331153Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8331802Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8332460Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8332601Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.8332686Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8332728Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8332767Z unimplemented []
2025-12-04T09:45:16.8332829Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8332929Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8333501Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8333540Z graph_break []
2025-12-04T09:45:16.8333613Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8333655Z Autotune Choices Stats:
2025-12-04T09:45:16.8334403Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.8334531Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8334647Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8334809Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8335424Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8336059Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8336666Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8337291Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8337891Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8338508Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8339129Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8339739Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8340361Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8340996Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8341139Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.8341181Z Autotune Choices Stats:
2025-12-04T09:45:16.8341963Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.8342181Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8342351Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8342632Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8343281Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8343931Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8344585Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8345212Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8345862Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8346492Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8347125Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8347761Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8348391Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8349036Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8349165Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.8349251Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8349293Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8349331Z unimplemented []
2025-12-04T09:45:16.8349392Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8349492Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8350083Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8350120Z graph_break []
2025-12-04T09:45:16.8350194Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8350234Z Autotune Choices Stats:
2025-12-04T09:45:16.8351007Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.8351136Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8351250Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8351413Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8352030Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8352640Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8353275Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8353902Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8354526Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8355132Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8355744Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8356361Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8356974Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8357597Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8357728Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.8357778Z Autotune Choices Stats:
2025-12-04T09:45:16.8358551Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.8358768Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8358935Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8359212Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8359848Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8360495Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8361122Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8361787Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8362416Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8363071Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8363700Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8364333Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8364965Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8365592Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8365734Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.8365809Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8365851Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8365889Z unimplemented []
2025-12-04T09:45:16.8365959Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8366059Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8366635Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8366685Z graph_break []
2025-12-04T09:45:16.8366769Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8366810Z Autotune Choices Stats:
2025-12-04T09:45:16.8367561Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.8367689Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8367805Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8367964Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8368584Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8369196Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8369806Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8370476Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8371112Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8371714Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8372328Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8372936Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8373643Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8374248Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8374390Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.8374433Z Autotune Choices Stats:
2025-12-04T09:45:16.8375207Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.8375435Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8375621Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8375896Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8376536Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8377177Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8377822Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8378455Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8379112Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8379753Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8380394Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8381056Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8381685Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8382316Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8382445Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.8382538Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8382581Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8382618Z unimplemented []
2025-12-04T09:45:16.8382681Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8382780Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8383377Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8383426Z graph_break []
2025-12-04T09:45:16.8383502Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8383542Z Autotune Choices Stats:
2025-12-04T09:45:16.8384293Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.8384421Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8384537Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8384696Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8385316Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8385928Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8386554Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8387174Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8387794Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8388420Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8389032Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8389641Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8390247Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8390885Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8391014Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.8391068Z Autotune Choices Stats:
2025-12-04T09:45:16.8391846Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.8392067Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8392251Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8392528Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8393170Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8393798Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8394424Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8395052Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8395681Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8396331Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8396980Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8397605Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8398237Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8398864Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8398992Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.8399068Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8399110Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8399148Z unimplemented []
2025-12-04T09:45:16.8399208Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8399309Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8399886Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8399935Z graph_break []
2025-12-04T09:45:16.8400008Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8400049Z Autotune Choices Stats:
2025-12-04T09:45:16.8400840Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.8400978Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8401105Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8401266Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8401888Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8402515Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8403124Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8403753Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8404370Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8404998Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8405626Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8406231Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8406841Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8407448Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8407578Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.8407620Z Autotune Choices Stats:
2025-12-04T09:45:16.8408381Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.8408613Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8408794Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8409074Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8409722Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8410347Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8411019Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8411640Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8412272Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8412916Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8413560Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8414208Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8414836Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8415481Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8415611Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.8415685Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8415728Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8415765Z unimplemented []
2025-12-04T09:45:16.8415825Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8415925Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8416503Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8416550Z graph_break []
2025-12-04T09:45:16.8416625Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8416664Z Autotune Choices Stats:
2025-12-04T09:45:16.8417416Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:16.8417544Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8417667Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8417829Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8418456Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8419065Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8419672Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8420288Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8420917Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8421539Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8422167Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8422793Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8423400Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8424008Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8424138Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:16.8424178Z Autotune Choices Stats:
2025-12-04T09:45:16.8424940Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:16.8425160Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8425346Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8425625Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8426265Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8426909Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8427535Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8428166Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8428800Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8429434Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8430074Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8430749Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8431400Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8432029Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8432159Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:16.8432235Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8432276Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8432315Z unimplemented []
2025-12-04T09:45:16.8432374Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8432475Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8433057Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8433096Z graph_break []
2025-12-04T09:45:16.8433169Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8433210Z Autotune Choices Stats:
2025-12-04T09:45:16.8433953Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.8434095Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8434211Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8434380Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8434997Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8435622Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8436229Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8436841Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8437445Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8438052Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8438679Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8439291Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8439918Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8440555Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8440686Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:16.8440727Z Autotune Choices Stats:
2025-12-04T09:45:16.8441500Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.8441720Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8441888Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8442166Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8442820Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8443463Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8444118Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8444743Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8445378Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8446011Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8446640Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8447289Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8447919Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8448578Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8448706Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:16.8448782Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8448825Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8448862Z unimplemented []
2025-12-04T09:45:16.8448923Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8449021Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8449595Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8449632Z graph_break []
2025-12-04T09:45:16.8449706Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8449745Z Autotune Choices Stats:
2025-12-04T09:45:16.8450524Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.8450652Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8450780Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8450944Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8451571Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8452179Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8452816Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8453424Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8454030Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8454636Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8455241Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8455872Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8456479Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8457115Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8457248Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:16.8457288Z Autotune Choices Stats:
2025-12-04T09:45:16.8458059Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.8458276Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8458444Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8458726Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8459375Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8460025Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8460684Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8461336Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8461968Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8462600Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8463231Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8463863Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8464520Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8465145Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8465296Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:16.8465370Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8465413Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8465450Z unimplemented []
2025-12-04T09:45:16.8465511Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8465611Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8466196Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8466235Z graph_break []
2025-12-04T09:45:16.8466310Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8466350Z Autotune Choices Stats:
2025-12-04T09:45:16.8467096Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:16.8467225Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8467341Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8467500Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8468120Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8468746Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8469363Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8469978Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8470617Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8471226Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8471840Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8472442Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8473085Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8473692Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8473846Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:16.8473887Z Autotune Choices Stats:
2025-12-04T09:45:16.8474648Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.8474868Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8475038Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8475317Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8475961Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8476606Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8477257Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8477880Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8478531Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8479162Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8479786Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8480452Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8481082Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8481737Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8481877Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:16.8481952Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8481995Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8482034Z unimplemented []
2025-12-04T09:45:16.8482093Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8482195Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8482783Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8482821Z graph_break []
2025-12-04T09:45:16.8482895Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8482934Z Autotune Choices Stats:
2025-12-04T09:45:16.8483688Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:16.8483816Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8483932Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8484096Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8484716Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8485320Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8485950Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8486571Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8487175Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8487791Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8488402Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8489020Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8489628Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8490258Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8490395Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:16.8490456Z Autotune Choices Stats:
2025-12-04T09:45:16.8491236Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.8491455Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8491621Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8491900Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8492526Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8493158Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8493786Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8494433Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8495081Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8495708Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8496334Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8496969Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8497609Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8498229Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8498371Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:16.8498444Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8498497Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8498535Z unimplemented []
2025-12-04T09:45:16.8498596Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8498695Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8499284Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8499333Z graph_break []
2025-12-04T09:45:16.8499408Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8499449Z Autotune Choices Stats:
2025-12-04T09:45:16.8500219Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:16.8500349Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8500497Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8500658Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8501277Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8501896Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8502518Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8503138Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8503778Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8504384Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8504996Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8505601Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8506222Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8506842Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8506973Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:16.8507013Z Autotune Choices Stats:
2025-12-04T09:45:16.8507787Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:16.8508030Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8508199Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8508486Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8509138Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8509761Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8510388Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8511034Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8511693Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8512347Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8512974Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8513608Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8514236Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8514865Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8515004Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:16.8515079Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8515121Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8515159Z unimplemented []
2025-12-04T09:45:16.8515220Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8515319Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8515905Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8515951Z graph_break []
2025-12-04T09:45:16.8516025Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8516064Z Autotune Choices Stats:
2025-12-04T09:45:16.8516820Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.8516948Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8517063Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8517225Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8517850Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8518475Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8519078Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8519697Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8520311Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8520982Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8521592Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8522203Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8522811Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8523418Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8523560Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:16.8523600Z Autotune Choices Stats:
2025-12-04T09:45:16.8524376Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.8524593Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8524774Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8525064Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8525699Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8526336Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8526967Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8527597Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8528235Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8528881Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8529525Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8530155Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8530822Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8531452Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8531588Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:16.8531661Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8531705Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8531741Z unimplemented []
2025-12-04T09:45:16.8531802Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8531903Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8532497Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8532539Z graph_break []
2025-12-04T09:45:16.8532613Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8532666Z Autotune Choices Stats:
2025-12-04T09:45:16.8533414Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.8533565Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8533681Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8533845Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8534461Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8535070Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8535684Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8536289Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8536904Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8537528Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8538157Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8538757Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8539366Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8539972Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8540103Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:16.8540143Z Autotune Choices Stats:
2025-12-04T09:45:16.8540941Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.8541174Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8541352Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8541641Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8542294Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8542922Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8543568Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8544211Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8544850Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8545492Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8546130Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8546781Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8547409Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8548036Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8548166Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:16.8548242Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8548284Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8548322Z unimplemented []
2025-12-04T09:45:16.8548382Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8548484Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8549063Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8549111Z graph_break []
2025-12-04T09:45:16.8549185Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8549224Z Autotune Choices Stats:
2025-12-04T09:45:16.8549983Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1708", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.8550119Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8550233Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8550396Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8551068Z   triton_flex_attention_1708 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8551676Z   triton_flex_attention_1706 0.0107 ms 93.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8552300Z   triton_flex_attention_1709 0.0110 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8552907Z   triton_flex_attention_1704 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8553533Z   triton_flex_attention_1707 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8554171Z   triton_flex_attention_1705 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8554778Z   triton_flex_attention_1724 0.0146 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8555409Z   triton_flex_attention_1716 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8556017Z   triton_flex_attention_1722 0.0160 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8556621Z   triton_flex_attention_1702 0.0166 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8556751Z SingleProcess AUTOTUNE benchmarking takes 0.1988 seconds and 0.5275 seconds precompiling for 24 choices
2025-12-04T09:45:16.8556791Z Autotune Choices Stats:
2025-12-04T09:45:16.8557560Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1743", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:16.8557780Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8557957Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8558237Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8558883Z   triton_flex_attention_backward_1743 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8559527Z   triton_flex_attention_backward_1737 0.0208 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8560153Z   triton_flex_attention_backward_1734 0.0213 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8560822Z   triton_flex_attention_backward_1735 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8561454Z   triton_flex_attention_backward_1745 0.0232 ms 76.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8562079Z   triton_flex_attention_backward_1744 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8562732Z   triton_flex_attention_backward_1742 0.0249 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8563362Z   triton_flex_attention_backward_1747 0.0252 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8564017Z   triton_flex_attention_backward_1738 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8564645Z   triton_flex_attention_backward_1729 0.0264 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8564775Z SingleProcess AUTOTUNE benchmarking takes 0.2428 seconds and 0.7372 seconds precompiling for 22 choices
2025-12-04T09:45:16.8564848Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8564891Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8564928Z unimplemented []
2025-12-04T09:45:16.8564989Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8565090Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8565662Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8565702Z graph_break []
2025-12-04T09:45:16.8565777Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8565818Z Autotune Choices Stats:
2025-12-04T09:45:16.8566566Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1754", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.8566705Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8566828Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8566990Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8567616Z   triton_flex_attention_1754 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8568239Z   triton_flex_attention_1752 0.0110 ms 90.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8568852Z   triton_flex_attention_1755 0.0114 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8569459Z   triton_flex_attention_1753 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8570069Z   triton_flex_attention_1750 0.0125 ms 79.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8570701Z   triton_flex_attention_1751 0.0143 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8571339Z   triton_flex_attention_1770 0.0149 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8571947Z   triton_flex_attention_1762 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8572572Z   triton_flex_attention_1768 0.0163 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8573181Z   triton_flex_attention_1748 0.0170 ms 58.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8573314Z SingleProcess AUTOTUNE benchmarking takes 0.2060 seconds and 0.4503 seconds precompiling for 24 choices
2025-12-04T09:45:16.8573356Z Autotune Choices Stats:
2025-12-04T09:45:16.8574145Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1789", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.8574369Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8574536Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8574827Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8575478Z   triton_flex_attention_backward_1789 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8576105Z   triton_flex_attention_backward_1783 0.0209 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8576750Z   triton_flex_attention_backward_1780 0.0216 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8577372Z   triton_flex_attention_backward_1781 0.0217 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8578010Z   triton_flex_attention_backward_1791 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8578640Z   triton_flex_attention_backward_1790 0.0235 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8579266Z   triton_flex_attention_backward_1788 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8579920Z   triton_flex_attention_backward_1793 0.0255 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8580578Z   triton_flex_attention_backward_1775 0.0264 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8581236Z   triton_flex_attention_backward_1784 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8581364Z SingleProcess AUTOTUNE benchmarking takes 0.2498 seconds and 0.6949 seconds precompiling for 22 choices
2025-12-04T09:45:16.8581439Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8581480Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8581518Z unimplemented []
2025-12-04T09:45:16.8581577Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8581679Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8582253Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8582292Z graph_break []
2025-12-04T09:45:16.8582366Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8582406Z Autotune Choices Stats:
2025-12-04T09:45:16.8583154Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1800", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.8583295Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8583411Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8583572Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8584202Z   triton_flex_attention_1800 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8584837Z   triton_flex_attention_1798 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8585445Z   triton_flex_attention_1801 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8586057Z   triton_flex_attention_1796 0.0121 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8586670Z   triton_flex_attention_1799 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8587282Z   triton_flex_attention_1816 0.0145 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8587889Z   triton_flex_attention_1797 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8588515Z   triton_flex_attention_1808 0.0152 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8589135Z   triton_flex_attention_1814 0.0161 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8589749Z   triton_flex_attention_1806 0.0168 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8589879Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.5450 seconds precompiling for 24 choices
2025-12-04T09:45:16.8589920Z Autotune Choices Stats:
2025-12-04T09:45:16.8590714Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1835", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.8590932Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8591101Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8591378Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8592012Z   triton_flex_attention_backward_1835 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8592668Z   triton_flex_attention_backward_1829 0.0210 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8593296Z   triton_flex_attention_backward_1826 0.0212 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8593948Z   triton_flex_attention_backward_1827 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8594581Z   triton_flex_attention_backward_1837 0.0231 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8595214Z   triton_flex_attention_backward_1836 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8595847Z   triton_flex_attention_backward_1839 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8596477Z   triton_flex_attention_backward_1834 0.0252 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8597123Z   triton_flex_attention_backward_1830 0.0260 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8597776Z   triton_flex_attention_backward_1821 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8597907Z SingleProcess AUTOTUNE benchmarking takes 0.2508 seconds and 0.7770 seconds precompiling for 22 choices
2025-12-04T09:45:16.8598001Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.8598049Z Traceback (most recent call last):
2025-12-04T09:45:16.8598204Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.8598245Z     self.assertTrue(
2025-12-04T09:45:16.8598355Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.8598403Z     raise self.failureException(msg)
2025-12-04T09:45:16.8598534Z AssertionError: False is not true : Log file /tmp/tmp2be0ko7i/flex_attention_configs.json was not created
2025-12-04T09:45:16.8598537Z 
2025-12-04T09:45:16.8598614Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.8598785Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.8598788Z 
2025-12-04T09:45:16.8598879Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.8598955Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8598997Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8599036Z unimplemented []
2025-12-04T09:45:16.8599098Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8599677Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.8599776Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8599814Z graph_break []
2025-12-04T09:45:16.8599888Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8600386Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.8600472Z   current_size = base.storage().size()
2025-12-04T09:45:16.8600515Z Autotune Choices Stats:
2025-12-04T09:45:16.8601288Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.8601430Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8601546Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8601709Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8602346Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8602952Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8603562Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8604164Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8604771Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8605396Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8606001Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8606629Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8607234Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8607835Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8607967Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.8608010Z Autotune Choices Stats:
2025-12-04T09:45:16.8608776Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.8609003Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8609172Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8609462Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8610093Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8610775Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8611397Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8612022Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8612647Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8613273Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8613921Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8614546Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8615194Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8615820Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8615951Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.8616028Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8616071Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8616108Z unimplemented []
2025-12-04T09:45:16.8616169Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8616270Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8616850Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8616885Z graph_break []
2025-12-04T09:45:16.8616961Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8617000Z Autotune Choices Stats:
2025-12-04T09:45:16.8617743Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.8617888Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8618012Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8618175Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8618806Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8619409Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8620014Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8620654Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8621254Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8621853Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8622489Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8623093Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8623721Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8624326Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8624456Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.8624496Z Autotune Choices Stats:
2025-12-04T09:45:16.8625258Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.8625480Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8625647Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8625938Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8626577Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8627205Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8627849Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8628473Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8629103Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8629727Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8630341Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8631029Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8631658Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8632314Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8632446Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.8632522Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8632564Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8632602Z unimplemented []
2025-12-04T09:45:16.8632663Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8632764Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8633340Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8633380Z graph_break []
2025-12-04T09:45:16.8633454Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8633496Z Autotune Choices Stats:
2025-12-04T09:45:16.8634245Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.8634386Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8634503Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8634665Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8635292Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8635908Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8636514Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8637117Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8637717Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8638328Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8638929Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8639554Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8640172Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8640829Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8640959Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.8641000Z Autotune Choices Stats:
2025-12-04T09:45:16.8641763Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.8641983Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8642151Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8642432Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8643061Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8643712Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8644332Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8644995Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8645625Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8646254Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8646885Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8647511Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8648155Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8648799Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8648939Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.8649017Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8649059Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8649095Z unimplemented []
2025-12-04T09:45:16.8649157Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8649257Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8649835Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8649871Z graph_break []
2025-12-04T09:45:16.8649946Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8649985Z Autotune Choices Stats:
2025-12-04T09:45:16.8650749Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.8650878Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8650993Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8651154Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8651786Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8652406Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8653036Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8653643Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8654253Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8654857Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8655489Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8656102Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8656738Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8657362Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8658390Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.8658430Z Autotune Choices Stats:
2025-12-04T09:45:16.8659192Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.8659416Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8659594Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8659877Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8660544Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8661164Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8661815Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8662457Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8663102Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8663768Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8664393Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8665027Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8665654Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8666302Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8666442Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.8666517Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8666561Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8666597Z unimplemented []
2025-12-04T09:45:16.8666660Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8666773Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8667364Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8667403Z graph_break []
2025-12-04T09:45:16.8667476Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8667517Z Autotune Choices Stats:
2025-12-04T09:45:16.8668267Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.8668400Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8668517Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8668679Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8669302Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8669913Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8670545Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8671178Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8671790Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8672389Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8672999Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8673607Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8674217Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8674828Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8674970Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.8675011Z Autotune Choices Stats:
2025-12-04T09:45:16.8675782Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.8676011Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8676177Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8676456Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8677091Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8677740Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8678357Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8678985Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8679633Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8680277Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8680934Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8681557Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8682187Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8682812Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8682942Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.8683032Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8683075Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8683114Z unimplemented []
2025-12-04T09:45:16.8683174Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8683290Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8683871Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8683924Z graph_break []
2025-12-04T09:45:16.8683998Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8684038Z Autotune Choices Stats:
2025-12-04T09:45:16.8684772Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.8684900Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8685016Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8685176Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8685788Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8686397Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8687006Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8687623Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8688246Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8688866Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8689487Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8690088Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8690739Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8691348Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8691477Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.8691522Z Autotune Choices Stats:
2025-12-04T09:45:16.8692300Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.8692542Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8692727Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8693010Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8693648Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8694278Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8694923Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8695548Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8696193Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8696862Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8697500Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8698135Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8698759Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8699383Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8699513Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.8699587Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8699629Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8699666Z unimplemented []
2025-12-04T09:45:16.8699727Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8699829Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8700457Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8700507Z graph_break []
2025-12-04T09:45:16.8700581Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8700621Z Autotune Choices Stats:
2025-12-04T09:45:16.8701378Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.8701522Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8701635Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8701797Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8702411Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8703013Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8703621Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8704230Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8704846Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8705469Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8706080Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8706695Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8707305Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8707909Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8708040Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.8708081Z Autotune Choices Stats:
2025-12-04T09:45:16.8708846Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.8709066Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8709243Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8709529Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8710181Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8710830Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8711453Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8712079Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8712710Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8713353Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8714014Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8714654Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8715284Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8715913Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8716043Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.8716121Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8716162Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8716200Z unimplemented []
2025-12-04T09:45:16.8716263Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8716365Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8716943Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8716982Z graph_break []
2025-12-04T09:45:16.8717054Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8717104Z Autotune Choices Stats:
2025-12-04T09:45:16.8717851Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.8717999Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8718127Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8718287Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8718898Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8719508Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8720116Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8720763Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8721363Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8721983Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8722613Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8723230Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8723836Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8724449Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8724579Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.8724620Z Autotune Choices Stats:
2025-12-04T09:45:16.8725380Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.8725599Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8725779Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8726056Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8726714Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8727345Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8727971Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8728595Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8729235Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8729866Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8730537Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8731190Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8731831Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8732456Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8732587Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.8732660Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8732702Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8732740Z unimplemented []
2025-12-04T09:45:16.8732802Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8732901Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8733480Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8733518Z graph_break []
2025-12-04T09:45:16.8733593Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8733634Z Autotune Choices Stats:
2025-12-04T09:45:16.8734388Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.8734526Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8734641Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8734804Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8735431Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8736044Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8736651Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8737257Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8737864Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8738479Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8739090Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8739717Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8740333Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8740977Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8741110Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.8741151Z Autotune Choices Stats:
2025-12-04T09:45:16.8741919Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.8742142Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8742310Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8742588Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8743239Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8743888Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8744533Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8745161Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8745808Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8746445Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8747084Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8747709Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8748360Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8748997Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8749127Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.8749202Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8749245Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8749285Z unimplemented []
2025-12-04T09:45:16.8749345Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8749447Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8750027Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8750064Z graph_break []
2025-12-04T09:45:16.8750139Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8750182Z Autotune Choices Stats:
2025-12-04T09:45:16.8750956Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.8751086Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8751222Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8751383Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8751996Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8752625Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8753246Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8753852Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8754457Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8755064Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8755684Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8756290Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8756916Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8757530Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8757661Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.8757706Z Autotune Choices Stats:
2025-12-04T09:45:16.8758465Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.8758684Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8758853Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8759129Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8759774Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8760432Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8761091Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8761725Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8762358Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8762990Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8763620Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8764266Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8764901Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8765543Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8765684Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.8765758Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8765802Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8765839Z unimplemented []
2025-12-04T09:45:16.8765900Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8766000Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8766577Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8766615Z graph_break []
2025-12-04T09:45:16.8766690Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8766730Z Autotune Choices Stats:
2025-12-04T09:45:16.8767487Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.8767617Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8767731Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8767891Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8768513Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8769132Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8769747Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8770362Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8771004Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8771611Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8772218Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8772856Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8773464Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8774092Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8774234Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.8774274Z Autotune Choices Stats:
2025-12-04T09:45:16.8775049Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.8775268Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8775436Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8775714Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8776351Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8776988Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8777608Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8778270Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8778908Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8779538Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8782318Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8782951Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8783601Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8784223Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8784387Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.8784477Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8784524Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8784563Z unimplemented []
2025-12-04T09:45:16.8784625Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8784728Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8785306Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8785346Z graph_break []
2025-12-04T09:45:16.8785421Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8785465Z Autotune Choices Stats:
2025-12-04T09:45:16.8786212Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.8786345Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8786465Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8786629Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8787265Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8787876Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8788500Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8789112Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8789716Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8790327Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8790982Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8791587Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8792205Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8792821Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8792964Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.8793017Z Autotune Choices Stats:
2025-12-04T09:45:16.8793783Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.8794004Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8794174Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8794450Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8795084Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8795707Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8796346Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8796964Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8797610Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8798251Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8798873Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8799505Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8800145Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8800831Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8800978Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.8801053Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8801096Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8801134Z unimplemented []
2025-12-04T09:45:16.8801195Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8801297Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8801882Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8801934Z graph_break []
2025-12-04T09:45:16.8802010Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8802051Z Autotune Choices Stats:
2025-12-04T09:45:16.8802790Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.8802919Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8803035Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8803199Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8803818Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8804443Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8805078Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8805702Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8806310Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8806920Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8807528Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8808136Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8808753Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8809380Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8809521Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.8809561Z Autotune Choices Stats:
2025-12-04T09:45:16.8810325Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.8810596Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8810763Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8811045Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8811700Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8812326Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8812951Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8813605Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8814257Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8814901Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8815522Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8816153Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8816778Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8817403Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8817534Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.8817608Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8817662Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8817702Z unimplemented []
2025-12-04T09:45:16.8817763Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8817863Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8818452Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8818500Z graph_break []
2025-12-04T09:45:16.8818583Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8818625Z Autotune Choices Stats:
2025-12-04T09:45:16.8819385Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.8819516Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8819632Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8819796Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8820448Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8821053Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8821664Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8822309Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8822946Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8823560Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8824163Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8824771Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8825393Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8826012Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8826146Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.8826187Z Autotune Choices Stats:
2025-12-04T09:45:16.8826949Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.8827192Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8827369Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8827648Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8828298Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8828918Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8829547Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8830188Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8830883Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8831540Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8832176Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8832805Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8833433Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8834077Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8834206Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.8834281Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8834323Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8834362Z unimplemented []
2025-12-04T09:45:16.8834421Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8834523Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8835109Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8835158Z graph_break []
2025-12-04T09:45:16.8835232Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8835271Z Autotune Choices Stats:
2025-12-04T09:45:16.8836035Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.8836170Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8836287Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8836448Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8837061Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8837668Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8838276Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8838875Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8839486Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8840114Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8840751Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8841371Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8841990Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8842612Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8842742Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.8842783Z Autotune Choices Stats:
2025-12-04T09:45:16.8843566Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.8843784Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8843969Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8844257Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8844903Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8845532Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8846170Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8846796Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8847424Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8848064Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8848708Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8849352Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8849982Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8850654Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8850784Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.8850858Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8850901Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8850938Z unimplemented []
2025-12-04T09:45:16.8850999Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8851099Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8851681Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8851720Z graph_break []
2025-12-04T09:45:16.8851794Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8851834Z Autotune Choices Stats:
2025-12-04T09:45:16.8852589Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.8852729Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8852857Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8853032Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8853649Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8854256Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8854871Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8855480Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8856086Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8856701Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8857328Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8857941Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8858552Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8859158Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8859287Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.8859327Z Autotune Choices Stats:
2025-12-04T09:45:16.8860089Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.8860310Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8860522Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8860799Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8861458Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8862097Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8862717Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8863340Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8863977Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8864631Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8865267Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8865909Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8866545Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8867173Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8867304Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.8867378Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8867419Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8867458Z unimplemented []
2025-12-04T09:45:16.8867518Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8867621Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8868201Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8868240Z graph_break []
2025-12-04T09:45:16.8868313Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8868355Z Autotune Choices Stats:
2025-12-04T09:45:16.8869102Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.8869230Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8869357Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8869519Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8870155Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8870812Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8871421Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8872033Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8872637Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8873247Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8873892Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8874525Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8875150Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8875757Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8875885Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.8875926Z Autotune Choices Stats:
2025-12-04T09:45:16.8876694Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.8876912Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8877080Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8877360Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8878011Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8878659Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8879294Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8879922Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8880621Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8881251Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8881890Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8882539Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8883193Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8883828Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8883958Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.8884035Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8884078Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8884115Z unimplemented []
2025-12-04T09:45:16.8884178Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8884277Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8884854Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8884891Z graph_break []
2025-12-04T09:45:16.8884968Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8885008Z Autotune Choices Stats:
2025-12-04T09:45:16.8885761Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.8885890Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8886005Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8886179Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8886794Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8887420Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8888034Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8888638Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8889242Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8889847Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8890489Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8891089Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8891725Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8892342Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8892474Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.8892515Z Autotune Choices Stats:
2025-12-04T09:45:16.8893270Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.8893490Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8893656Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8893937Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8894572Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8895211Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8895856Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8896497Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8897131Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8897779Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8898406Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8899049Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8899676Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8900319Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8900495Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.8900572Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8900613Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8900651Z unimplemented []
2025-12-04T09:45:16.8900711Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8900812Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8901390Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8901428Z graph_break []
2025-12-04T09:45:16.8901501Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8901542Z Autotune Choices Stats:
2025-12-04T09:45:16.8902289Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.8902416Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8902532Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8902692Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8903328Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8903932Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8904563Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8905183Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8905797Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8906405Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8907013Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8907632Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8908240Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8908878Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8909018Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.8909062Z Autotune Choices Stats:
2025-12-04T09:45:16.8909824Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.8910041Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8910208Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8910517Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8911148Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8911789Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8912415Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8913066Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8913707Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8914343Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8914968Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8915597Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8916231Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8916861Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8917009Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.8917094Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8917136Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8917174Z unimplemented []
2025-12-04T09:45:16.8917235Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8917334Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8917912Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8917948Z graph_break []
2025-12-04T09:45:16.8918023Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8918064Z Autotune Choices Stats:
2025-12-04T09:45:16.8918809Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.8918937Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8919052Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8919216Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8919829Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8920483Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8921111Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8921727Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8922335Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8922942Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8923547Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8924152Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8924777Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8925382Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8925531Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.8925581Z Autotune Choices Stats:
2025-12-04T09:45:16.8926341Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.8926560Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8926726Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8927007Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8927645Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8928273Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8928910Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8929536Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8930186Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8930844Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8931466Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8932103Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8932730Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8933375Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8933517Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.8933591Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8933633Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8933671Z unimplemented []
2025-12-04T09:45:16.8933730Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8933830Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8934419Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.8934469Z graph_break []
2025-12-04T09:45:16.8934541Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8934583Z Autotune Choices Stats:
2025-12-04T09:45:16.8935328Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.8935455Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8935571Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8935732Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8936359Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8936959Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8937589Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8938213Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8938839Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8939444Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8940051Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8940690Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8941294Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8941918Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8942060Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.8942103Z Autotune Choices Stats:
2025-12-04T09:45:16.8942868Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.8943100Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8943272Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8943554Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8944191Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8944817Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8945466Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8946108Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8946733Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8947379Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8948017Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8948645Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8949275Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8949903Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8950034Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.8950109Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8950153Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8950200Z unimplemented []
2025-12-04T09:45:16.8950262Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8950362Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8950987Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8951023Z graph_break []
2025-12-04T09:45:16.8951113Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8951169Z Autotune Choices Stats:
2025-12-04T09:45:16.8951916Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.8952046Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8952163Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8952326Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8952941Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8953544Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8954153Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8954773Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8955403Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8956010Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8956621Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8957226Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8957832Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8958438Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8958569Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.8958611Z Autotune Choices Stats:
2025-12-04T09:45:16.8959389Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.8959616Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8959792Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8960080Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8960758Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8961389Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8962011Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8962636Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8963282Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8963934Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8964572Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8965207Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8965838Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8966469Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8966600Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.8966674Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8966717Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8966755Z unimplemented []
2025-12-04T09:45:16.8966815Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8966917Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8967504Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8967554Z graph_break []
2025-12-04T09:45:16.8967627Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8967668Z Autotune Choices Stats:
2025-12-04T09:45:16.8968417Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.8968559Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8968676Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8968836Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8969461Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8970061Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8970689Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8971295Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8971920Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8972558Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8973180Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8973809Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8974434Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8975058Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8975188Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.8975229Z Autotune Choices Stats:
2025-12-04T09:45:16.8976010Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.8976230Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8976405Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8976694Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8977337Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8977969Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8978595Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8979222Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8979869Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8980554Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8981203Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8981848Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8982475Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8983106Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8983235Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.8983311Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8983352Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8983391Z unimplemented []
2025-12-04T09:45:16.8983452Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.8983552Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.8984129Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.8984165Z graph_break []
2025-12-04T09:45:16.8984239Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.8984279Z Autotune Choices Stats:
2025-12-04T09:45:16.8985033Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.8985171Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8985293Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8985469Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8986088Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8986696Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8987309Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8987910Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8988509Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8989128Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8989763Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8990380Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8991022Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8991629Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8991760Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.8991801Z Autotune Choices Stats:
2025-12-04T09:45:16.8992561Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.8992780Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.8992962Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.8993238Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.8993891Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8994530Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8995156Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8995784Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8996412Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8997045Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8997697Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8998350Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.8998992Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8999620Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.8999751Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.8999825Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.8999867Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.8999904Z unimplemented []
2025-12-04T09:45:16.8999964Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9000062Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9000686Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9000724Z graph_break []
2025-12-04T09:45:16.9000796Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9000837Z Autotune Choices Stats:
2025-12-04T09:45:16.9001603Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.9001732Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9001861Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9002024Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9002652Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9003276Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9003884Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9004494Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9005104Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9005724Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9006339Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9006967Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9007588Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9008199Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9008329Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.9008370Z Autotune Choices Stats:
2025-12-04T09:45:16.9009137Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.9009355Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9009522Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9009803Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9010484Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9011140Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9011776Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9012404Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9013040Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9013677Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9014301Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9014940Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9015585Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9016222Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9016351Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.9016428Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9016471Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9016509Z unimplemented []
2025-12-04T09:45:16.9016569Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9016671Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9017250Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9017287Z graph_break []
2025-12-04T09:45:16.9017364Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9017403Z Autotune Choices Stats:
2025-12-04T09:45:16.9018143Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.9018271Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9018386Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9018560Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9019168Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9019803Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9020459Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9021065Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9021670Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9022277Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9022907Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9023514Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9024150Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9024766Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9024895Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.9024935Z Autotune Choices Stats:
2025-12-04T09:45:16.9025704Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.9025922Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9026092Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9026373Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9027008Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9027640Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9028284Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9028921Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9029559Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9030186Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9030847Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9031492Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9032118Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9032774Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9032915Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.9032990Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9033032Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9033069Z unimplemented []
2025-12-04T09:45:16.9033131Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9033232Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9033808Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9033847Z graph_break []
2025-12-04T09:45:16.9033921Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9033965Z Autotune Choices Stats:
2025-12-04T09:45:16.9034730Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.9034860Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9034975Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9035139Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9035769Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9036391Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9037010Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9037623Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9038235Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9038835Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9039438Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9040055Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9040699Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9041332Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9041478Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.9041520Z Autotune Choices Stats:
2025-12-04T09:45:16.9042282Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.9042504Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9042671Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9042950Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9043590Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9044231Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9044858Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9045506Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9046149Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9046781Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9047405Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9048032Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9048672Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9049301Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9049451Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.9049536Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9049579Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9049617Z unimplemented []
2025-12-04T09:45:16.9049677Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9049776Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9050356Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9050394Z graph_break []
2025-12-04T09:45:16.9050503Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9050543Z Autotune Choices Stats:
2025-12-04T09:45:16.9051292Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.9051421Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9051538Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9051700Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9052318Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9052949Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9053586Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9054202Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9054806Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9055413Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9056027Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9056632Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9057243Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9057861Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9058002Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.9058055Z Autotune Choices Stats:
2025-12-04T09:45:16.9058819Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.9059037Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9059205Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9059493Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9060148Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9060805Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9061449Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9062090Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9062733Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9063370Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9063999Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9064627Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9065260Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9065900Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9066039Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.9066115Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9066158Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9066194Z unimplemented []
2025-12-04T09:45:16.9066256Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9066367Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9066960Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9066998Z graph_break []
2025-12-04T09:45:16.9067070Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9067111Z Autotune Choices Stats:
2025-12-04T09:45:16.9067859Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.9067987Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9068101Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9068264Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9068885Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9069492Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9070112Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9070780Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9071397Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9072004Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9072611Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9073234Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9073851Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9074472Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9074615Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.9074656Z Autotune Choices Stats:
2025-12-04T09:45:16.9075436Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.9075667Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9075833Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9076114Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9076752Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9077402Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9078036Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9078665Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9079313Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9079958Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9080627Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9081251Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9081883Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9082513Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9082640Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.9082732Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9082774Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9082812Z unimplemented []
2025-12-04T09:45:16.9082873Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9082987Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9083578Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9083629Z graph_break []
2025-12-04T09:45:16.9083704Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9083744Z Autotune Choices Stats:
2025-12-04T09:45:16.9084496Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:16.9084623Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9084738Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9084902Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9085521Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9086131Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9086745Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9087364Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9087990Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9088603Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9089215Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9089825Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9090472Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9091076Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9091205Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:16.9091260Z Autotune Choices Stats:
2025-12-04T09:45:16.9092024Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:16.9092265Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9092445Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9092727Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9093367Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9093992Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9094621Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9095249Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9095883Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9096528Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9097162Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9097798Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9098426Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9099065Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9099197Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:16.9099270Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9099313Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9099349Z unimplemented []
2025-12-04T09:45:16.9099411Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9099511Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9100099Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9100151Z graph_break []
2025-12-04T09:45:16.9100224Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9100264Z Autotune Choices Stats:
2025-12-04T09:45:16.9101045Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.9101190Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9101303Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9101463Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9102078Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9102687Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9103314Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9103939Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9104563Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9105192Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9105813Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9106422Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9107029Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9107637Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9107769Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:16.9107810Z Autotune Choices Stats:
2025-12-04T09:45:16.9108584Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.9108813Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9108977Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9109269Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9109917Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9110578Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9111198Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9111827Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9112464Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9113112Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9113755Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9114399Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9115028Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9115660Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9115788Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:16.9115863Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9115905Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9115943Z unimplemented []
2025-12-04T09:45:16.9116004Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9116105Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9116701Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9116738Z graph_break []
2025-12-04T09:45:16.9116828Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9116870Z Autotune Choices Stats:
2025-12-04T09:45:16.9117624Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.9117771Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9117897Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9118056Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9118675Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9119286Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9119893Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9120535Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9121156Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9121763Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9122396Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9123013Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9123622Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9124230Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9124359Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:16.9124400Z Autotune Choices Stats:
2025-12-04T09:45:16.9125160Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.9125377Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9125557Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9125848Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9126493Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9127131Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9127764Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9128390Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9129030Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9129674Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9130307Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9131003Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9131645Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9132278Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9132409Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:16.9132485Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9132528Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9132564Z unimplemented []
2025-12-04T09:45:16.9132625Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9132726Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9133308Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9133345Z graph_break []
2025-12-04T09:45:16.9133424Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9133463Z Autotune Choices Stats:
2025-12-04T09:45:16.9134226Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:16.9134366Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9134480Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9134642Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9135275Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9135890Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9136494Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9137102Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9137711Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9138328Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9138939Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9139563Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9140187Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9140806Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9140936Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:16.9140976Z Autotune Choices Stats:
2025-12-04T09:45:16.9141742Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.9141959Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9142126Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9142419Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9143055Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9143708Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9144355Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9145000Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9145628Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9146261Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9146900Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9147532Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9148194Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9148835Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9148964Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:16.9149042Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9149083Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9149121Z unimplemented []
2025-12-04T09:45:16.9149182Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9149283Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9149861Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9149898Z graph_break []
2025-12-04T09:45:16.9149972Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9150013Z Autotune Choices Stats:
2025-12-04T09:45:16.9150793Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:16.9150921Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9151053Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9151214Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9151854Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9152472Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9153080Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9153687Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9154296Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9154901Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9155521Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9156141Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9156763Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9157380Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9157509Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:16.9157553Z Autotune Choices Stats:
2025-12-04T09:45:16.9158312Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.9158533Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9158703Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9158982Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9159638Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9160266Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9160945Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9161577Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9162214Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9162847Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9163479Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9164120Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9164769Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9165418Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9165561Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:16.9165637Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9165680Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9165719Z unimplemented []
2025-12-04T09:45:16.9165781Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9165880Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9166459Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9166496Z graph_break []
2025-12-04T09:45:16.9166570Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9166609Z Autotune Choices Stats:
2025-12-04T09:45:16.9167351Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:16.9167480Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9167594Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9167757Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9168387Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9169014Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9169635Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9170236Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9170883Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9171489Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9172103Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9172719Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9173344Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9173960Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9174091Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:16.9174130Z Autotune Choices Stats:
2025-12-04T09:45:16.9174897Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:16.9175117Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9175284Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9175562Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9176197Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9176826Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9177464Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9178101Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9178743Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9179374Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9180004Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9180674Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9181326Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9181978Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9182121Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:16.9182194Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9182234Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9182272Z unimplemented []
2025-12-04T09:45:16.9182335Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9182435Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9183012Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9183052Z graph_break []
2025-12-04T09:45:16.9183126Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9183168Z Autotune Choices Stats:
2025-12-04T09:45:16.9183921Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.9184050Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9184167Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9184327Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9184945Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9185564Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9186188Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9186805Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9187410Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9188015Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9188628Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9189234Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9189859Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9190523Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9190665Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:16.9190707Z Autotune Choices Stats:
2025-12-04T09:45:16.9191466Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.9191687Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9191854Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9192136Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9192770Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9193397Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9194056Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9194707Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9195355Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9195988Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9196614Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9197251Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9197894Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9198532Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9198671Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:16.9198746Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9198789Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9198826Z unimplemented []
2025-12-04T09:45:16.9198899Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9199009Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9199587Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9199624Z graph_break []
2025-12-04T09:45:16.9199700Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9199739Z Autotune Choices Stats:
2025-12-04T09:45:16.9200547Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.9200676Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9200788Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9200952Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9201569Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9202178Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9202806Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9203428Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9204042Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9204649Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9205261Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9205891Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9206519Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9207142Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9207281Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:16.9207321Z Autotune Choices Stats:
2025-12-04T09:45:16.9208095Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.9208325Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9208491Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9208773Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9209415Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9210057Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9210728Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9211363Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9212028Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9212679Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9213310Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9213935Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9214571Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9215198Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9215339Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:16.9215415Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9215457Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9215509Z unimplemented []
2025-12-04T09:45:16.9215569Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9215671Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9216256Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9216304Z graph_break []
2025-12-04T09:45:16.9216378Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9216419Z Autotune Choices Stats:
2025-12-04T09:45:16.9217168Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1708", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.9217297Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9217414Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9217575Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9218200Z   triton_flex_attention_1708 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9218819Z   triton_flex_attention_1706 0.0107 ms 93.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9219442Z   triton_flex_attention_1709 0.0110 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9220050Z   triton_flex_attention_1704 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9220710Z   triton_flex_attention_1707 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9221331Z   triton_flex_attention_1705 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9221947Z   triton_flex_attention_1724 0.0146 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9222555Z   triton_flex_attention_1716 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9223163Z   triton_flex_attention_1722 0.0160 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9223764Z   triton_flex_attention_1702 0.0166 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9223906Z SingleProcess AUTOTUNE benchmarking takes 0.1988 seconds and 0.5275 seconds precompiling for 24 choices
2025-12-04T09:45:16.9223947Z Autotune Choices Stats:
2025-12-04T09:45:16.9224717Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1743", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:16.9224958Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9225124Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9225406Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9226037Z   triton_flex_attention_backward_1743 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9226664Z   triton_flex_attention_backward_1737 0.0208 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9227299Z   triton_flex_attention_backward_1734 0.0213 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9227931Z   triton_flex_attention_backward_1735 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9228564Z   triton_flex_attention_backward_1745 0.0232 ms 76.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9229221Z   triton_flex_attention_backward_1744 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9229853Z   triton_flex_attention_backward_1742 0.0249 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9230510Z   triton_flex_attention_backward_1747 0.0252 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9231145Z   triton_flex_attention_backward_1738 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9231772Z   triton_flex_attention_backward_1729 0.0264 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9231904Z SingleProcess AUTOTUNE benchmarking takes 0.2428 seconds and 0.7372 seconds precompiling for 22 choices
2025-12-04T09:45:16.9231981Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9232024Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9232062Z unimplemented []
2025-12-04T09:45:16.9232125Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9232226Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9232821Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9232870Z graph_break []
2025-12-04T09:45:16.9232944Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9232984Z Autotune Choices Stats:
2025-12-04T09:45:16.9233747Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1754", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.9233888Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9234001Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9234163Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9234773Z   triton_flex_attention_1754 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9235380Z   triton_flex_attention_1752 0.0110 ms 90.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9235992Z   triton_flex_attention_1755 0.0114 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9236608Z   triton_flex_attention_1753 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9237216Z   triton_flex_attention_1750 0.0125 ms 79.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9237845Z   triton_flex_attention_1751 0.0143 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9238465Z   triton_flex_attention_1770 0.0149 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9239067Z   triton_flex_attention_1762 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9239675Z   triton_flex_attention_1768 0.0163 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9240281Z   triton_flex_attention_1748 0.0170 ms 58.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9240442Z SingleProcess AUTOTUNE benchmarking takes 0.2060 seconds and 0.4503 seconds precompiling for 24 choices
2025-12-04T09:45:16.9240484Z Autotune Choices Stats:
2025-12-04T09:45:16.9241266Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1789", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.9241499Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9241676Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9241970Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9242607Z   triton_flex_attention_backward_1789 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9243233Z   triton_flex_attention_backward_1783 0.0209 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9243859Z   triton_flex_attention_backward_1780 0.0216 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9244503Z   triton_flex_attention_backward_1781 0.0217 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9245159Z   triton_flex_attention_backward_1791 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9245786Z   triton_flex_attention_backward_1790 0.0235 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9246432Z   triton_flex_attention_backward_1788 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9247068Z   triton_flex_attention_backward_1793 0.0255 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9247697Z   triton_flex_attention_backward_1775 0.0264 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9248327Z   triton_flex_attention_backward_1784 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9248459Z SingleProcess AUTOTUNE benchmarking takes 0.2498 seconds and 0.6949 seconds precompiling for 22 choices
2025-12-04T09:45:16.9248534Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9248577Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9248617Z unimplemented []
2025-12-04T09:45:16.9248679Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9248781Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9249363Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9249403Z graph_break []
2025-12-04T09:45:16.9249476Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9249518Z Autotune Choices Stats:
2025-12-04T09:45:16.9250283Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1800", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.9250451Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9250566Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9250728Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9251351Z   triton_flex_attention_1800 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9251978Z   triton_flex_attention_1798 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9252605Z   triton_flex_attention_1801 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9253205Z   triton_flex_attention_1796 0.0121 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9253825Z   triton_flex_attention_1799 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9254435Z   triton_flex_attention_1816 0.0145 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9255063Z   triton_flex_attention_1797 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9255680Z   triton_flex_attention_1808 0.0152 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9256290Z   triton_flex_attention_1814 0.0161 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9256899Z   triton_flex_attention_1806 0.0168 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9257028Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.5450 seconds precompiling for 24 choices
2025-12-04T09:45:16.9257069Z Autotune Choices Stats:
2025-12-04T09:45:16.9257833Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1835", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.9258063Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9258230Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9258518Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9259166Z   triton_flex_attention_backward_1835 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9259811Z   triton_flex_attention_backward_1829 0.0210 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9260477Z   triton_flex_attention_backward_1826 0.0212 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9261109Z   triton_flex_attention_backward_1827 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9261746Z   triton_flex_attention_backward_1837 0.0231 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9262394Z   triton_flex_attention_backward_1836 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9263024Z   triton_flex_attention_backward_1839 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9263676Z   triton_flex_attention_backward_1834 0.0252 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9264310Z   triton_flex_attention_backward_1830 0.0260 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9264941Z   triton_flex_attention_backward_1821 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9265069Z SingleProcess AUTOTUNE benchmarking takes 0.2508 seconds and 0.7770 seconds precompiling for 22 choices
2025-12-04T09:45:16.9265145Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9265186Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9265225Z unimplemented []
2025-12-04T09:45:16.9265287Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9265389Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9265968Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9266006Z graph_break []
2025-12-04T09:45:16.9266081Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9266121Z Autotune Choices Stats:
2025-12-04T09:45:16.9266882Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1846", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:16.9267020Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9267133Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9267305Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9267931Z   triton_flex_attention_1846 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9268533Z   triton_flex_attention_1844 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9269140Z   triton_flex_attention_1847 0.0114 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9269746Z   triton_flex_attention_1842 0.0122 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9270347Z   triton_flex_attention_1845 0.0124 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9271071Z   triton_flex_attention_1843 0.0144 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9271686Z   triton_flex_attention_1862 0.0146 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9272300Z   triton_flex_attention_1854 0.0154 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9272919Z   triton_flex_attention_1860 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9273527Z   triton_flex_attention_1840 0.0167 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9273656Z SingleProcess AUTOTUNE benchmarking takes 0.2278 seconds and 0.3492 seconds precompiling for 24 choices
2025-12-04T09:45:16.9273697Z Autotune Choices Stats:
2025-12-04T09:45:16.9274460Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1881", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.9274681Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9274848Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9275141Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9275775Z   triton_flex_attention_backward_1881 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9276427Z   triton_flex_attention_backward_1875 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9277066Z   triton_flex_attention_backward_1873 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9277692Z   triton_flex_attention_backward_1872 0.0216 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9278319Z   triton_flex_attention_backward_1882 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9278954Z   triton_flex_attention_backward_1883 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9279594Z   triton_flex_attention_backward_1880 0.0254 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9280232Z   triton_flex_attention_backward_1885 0.0254 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9280931Z   triton_flex_attention_backward_1876 0.0263 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9281570Z   triton_flex_attention_backward_1867 0.0267 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9281700Z SingleProcess AUTOTUNE benchmarking takes 0.2409 seconds and 0.8665 seconds precompiling for 22 choices
2025-12-04T09:45:16.9281794Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:16.9281845Z Traceback (most recent call last):
2025-12-04T09:45:16.9281999Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:16.9282040Z     self.assertTrue(
2025-12-04T09:45:16.9282147Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:16.9282196Z     raise self.failureException(msg)
2025-12-04T09:45:16.9282324Z AssertionError: False is not true : Log file /tmp/tmpop9htqnm/flex_attention_configs.json was not created
2025-12-04T09:45:16.9282327Z 
2025-12-04T09:45:16.9282403Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:16.9282572Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:16.9282575Z 
2025-12-04T09:45:16.9282663Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:16.9282738Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9282782Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9282820Z unimplemented []
2025-12-04T09:45:16.9282880Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9283461Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:16.9283572Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9283610Z graph_break []
2025-12-04T09:45:16.9283683Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9284189Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:16.9284238Z   current_size = base.storage().size()
2025-12-04T09:45:16.9284278Z Autotune Choices Stats:
2025-12-04T09:45:16.9285032Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.9285172Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9285287Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9287056Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9287675Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9288294Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9288903Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9289523Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9290126Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9290794Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9291422Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9292027Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9292637Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9293243Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9293375Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:16.9293415Z Autotune Choices Stats:
2025-12-04T09:45:16.9294199Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.9294440Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9294620Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9294910Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9295556Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9296177Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9296800Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9297437Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9298072Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9298697Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9299339Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9299981Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9300647Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9301270Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9301401Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:16.9301476Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9301520Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9301559Z unimplemented []
2025-12-04T09:45:16.9301620Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9301721Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9302324Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9302364Z graph_break []
2025-12-04T09:45:16.9302437Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9302493Z Autotune Choices Stats:
2025-12-04T09:45:16.9303254Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.9303398Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9303513Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9303677Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9304288Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9304894Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9305499Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9306099Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9306707Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9307318Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9307935Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9308548Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9309153Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9309753Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9309886Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:16.9309927Z Autotune Choices Stats:
2025-12-04T09:45:16.9310721Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.9310971Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9311138Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9311441Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9312087Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9312724Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9313346Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9313966Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9314596Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9315237Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9315858Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9316512Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9317152Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9317776Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9317907Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:16.9317982Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9318023Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9318062Z unimplemented []
2025-12-04T09:45:16.9318123Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9318225Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9318801Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9318839Z graph_break []
2025-12-04T09:45:16.9318913Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9318953Z Autotune Choices Stats:
2025-12-04T09:45:16.9319708Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:16.9319846Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9319960Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9320136Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9320791Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9321398Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9322008Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9322618Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9323223Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9323859Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9324470Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9325083Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9325704Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9326309Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9326441Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:16.9326482Z Autotune Choices Stats:
2025-12-04T09:45:16.9327244Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.9327462Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9327630Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9327923Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9328557Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9329206Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9329844Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9330518Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9331155Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9331788Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9332440Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9333081Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9333716Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9334350Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9334596Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:16.9334691Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9334743Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9334781Z unimplemented []
2025-12-04T09:45:16.9334842Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9334942Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9335516Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9335555Z graph_break []
2025-12-04T09:45:16.9335630Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9335669Z Autotune Choices Stats:
2025-12-04T09:45:16.9336434Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:16.9336582Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9336696Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9336868Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9337489Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9338102Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9338713Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9339321Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9339930Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9340566Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9341199Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9341842Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9342460Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9343062Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9343192Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:16.9343231Z Autotune Choices Stats:
2025-12-04T09:45:16.9343985Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.9344209Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9344377Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9344654Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9345294Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9345931Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9346569Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9347202Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9347829Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9348456Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9349082Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9349739Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9350387Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9351042Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9351172Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:16.9351247Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9351288Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9351327Z unimplemented []
2025-12-04T09:45:16.9351388Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9351489Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9352078Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9352116Z graph_break []
2025-12-04T09:45:16.9352189Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9352230Z Autotune Choices Stats:
2025-12-04T09:45:16.9352975Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:16.9353102Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9353217Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9353378Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9354017Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9354647Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9355266Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9355884Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9356507Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9357116Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9357724Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9358341Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9358973Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9359584Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9359713Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:16.9359755Z Autotune Choices Stats:
2025-12-04T09:45:16.9360554Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:16.9360774Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9360942Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9361221Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9361859Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9362508Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9363163Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9363795Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9364427Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9365055Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9365680Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9366305Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9366945Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9367591Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9367730Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:16.9367803Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9367846Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9367884Z unimplemented []
2025-12-04T09:45:16.9367946Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9368044Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9368618Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9368656Z graph_break []
2025-12-04T09:45:16.9368730Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9368769Z Autotune Choices Stats:
2025-12-04T09:45:16.9369506Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:16.9369636Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9369750Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9369911Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9370579Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9371208Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9371836Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9372452Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9373051Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9373652Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9374261Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9374867Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9375487Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9376111Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9376251Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:16.9376291Z Autotune Choices Stats:
2025-12-04T09:45:16.9377055Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.9377276Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9377440Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9377720Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9378354Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9378981Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9379614Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9380253Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9380921Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9381550Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9382171Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9382798Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9383420Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9384067Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9384208Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:16.9384283Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9384323Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9384363Z unimplemented []
2025-12-04T09:45:16.9384439Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9384557Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9385134Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9385172Z graph_break []
2025-12-04T09:45:16.9385245Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9385286Z Autotune Choices Stats:
2025-12-04T09:45:16.9386033Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:16.9386161Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9386275Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9386436Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9387052Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9387658Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9388275Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9388896Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9389512Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9390115Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9390761Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9391369Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9391977Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9392600Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9392740Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:16.9392782Z Autotune Choices Stats:
2025-12-04T09:45:16.9393576Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:16.9393806Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9393973Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9394256Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9394888Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9395513Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9396153Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9396789Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9397439Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9398075Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9398716Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9399343Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9399971Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9400641Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9400791Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:16.9400864Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9400907Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9400959Z unimplemented []
2025-12-04T09:45:16.9401020Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9401120Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9401715Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9401764Z graph_break []
2025-12-04T09:45:16.9401838Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9401878Z Autotune Choices Stats:
2025-12-04T09:45:16.9402619Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:16.9402748Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9402862Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9403027Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9403639Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9404256Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9404872Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9405477Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9406103Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9406714Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9407319Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9407942Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9408546Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9409154Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9409294Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:16.9409334Z Autotune Choices Stats:
2025-12-04T09:45:16.9410099Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.9410338Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9410536Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9410816Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9411469Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9412111Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9412756Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9413391Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9414041Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9414703Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9415332Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9415961Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9416594Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9417217Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9417347Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:16.9417421Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9417463Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9417502Z unimplemented []
2025-12-04T09:45:16.9417564Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9417664Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9418254Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9418302Z graph_break []
2025-12-04T09:45:16.9418375Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9418415Z Autotune Choices Stats:
2025-12-04T09:45:16.9419172Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.9419311Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9419425Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9419585Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9420202Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9420854Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9421458Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9422087Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9422694Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9423326Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9423945Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9424547Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9425151Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9425752Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9425880Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:16.9425920Z Autotune Choices Stats:
2025-12-04T09:45:16.9426682Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.9426913Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9427089Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9427378Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9428010Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9428640Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9429272Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9429895Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9430537Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9431183Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9431834Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9432471Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9433098Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9433723Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9433853Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:16.9433925Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9433968Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9434006Z unimplemented []
2025-12-04T09:45:16.9434067Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9434167Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9434744Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9434793Z graph_break []
2025-12-04T09:45:16.9434869Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9434908Z Autotune Choices Stats:
2025-12-04T09:45:16.9435659Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:16.9435807Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9435921Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9436082Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9436712Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9437323Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9437931Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9438556Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9439180Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9439780Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9440437Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9441055Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9441652Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9442257Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9442387Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:16.9442426Z Autotune Choices Stats:
2025-12-04T09:45:16.9443185Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.9443429Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9443595Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9443890Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9444541Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9445177Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9445814Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9446442Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9447079Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9447725Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9448340Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9448987Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9449625Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9450263Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9450393Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:16.9450503Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9450546Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9450584Z unimplemented []
2025-12-04T09:45:16.9450647Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9450749Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9451333Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9451370Z graph_break []
2025-12-04T09:45:16.9451443Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9451483Z Autotune Choices Stats:
2025-12-04T09:45:16.9452264Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:16.9452406Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9452521Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9452692Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9453329Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9453940Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9454554Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9455164Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9455787Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9456417Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9457021Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9457648Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9458263Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9458881Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9459009Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:16.9459049Z Autotune Choices Stats:
2025-12-04T09:45:16.9459812Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.9460032Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9460200Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9460518Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9461156Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9461809Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9462445Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9463072Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9463704Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9464330Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9464966Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9465591Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9466234Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9466871Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9467001Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:16.9467075Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9467117Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9467156Z unimplemented []
2025-12-04T09:45:16.9467216Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9467316Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9467892Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9467930Z graph_break []
2025-12-04T09:45:16.9468004Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9468044Z Autotune Choices Stats:
2025-12-04T09:45:16.9468789Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:16.9468920Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9469045Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9469207Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9469848Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9470499Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9471109Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9471731Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9472336Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9472938Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9473568Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9474186Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9474815Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9475429Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9475559Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:16.9475598Z Autotune Choices Stats:
2025-12-04T09:45:16.9476359Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.9476581Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9476747Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9477028Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9477670Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9478306Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9478949Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9479583Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9480230Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9480908Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9481535Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9482181Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9482822Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9483456Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9483597Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:16.9483670Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9483712Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9483751Z unimplemented []
2025-12-04T09:45:16.9483812Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9483912Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9484487Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9484526Z graph_break []
2025-12-04T09:45:16.9484600Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9484641Z Autotune Choices Stats:
2025-12-04T09:45:16.9485383Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.9485512Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9485625Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9485787Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9486410Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9487048Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9487665Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9488270Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9488874Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9489481Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9490088Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9490763Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9491391Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9492005Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9492138Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:16.9492179Z Autotune Choices Stats:
2025-12-04T09:45:16.9492945Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:16.9493164Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9493330Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9493607Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9494242Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9494882Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9495513Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9496141Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9496777Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9497405Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9498028Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9498658Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9499300Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9499948Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9500087Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:16.9500163Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9500204Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9500244Z unimplemented []
2025-12-04T09:45:16.9500306Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9500439Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9501014Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9501051Z graph_break []
2025-12-04T09:45:16.9501128Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9501167Z Autotune Choices Stats:
2025-12-04T09:45:16.9501914Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:16.9502046Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9502164Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9502324Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9502939Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9503585Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9504224Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9504839Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9505443Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9506062Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9506684Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9507290Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9507904Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9508525Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9508664Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:16.9508703Z Autotune Choices Stats:
2025-12-04T09:45:16.9509456Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.9509675Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9509843Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9510121Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9510782Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9511405Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9512065Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9512719Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9513356Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9513982Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9514604Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9515228Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9515855Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9516491Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9516630Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:16.9516704Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9516747Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9516785Z unimplemented []
2025-12-04T09:45:16.9516847Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9516957Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9517552Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9517590Z graph_break []
2025-12-04T09:45:16.9517663Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9517704Z Autotune Choices Stats:
2025-12-04T09:45:16.9518441Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:16.9518569Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9518682Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9518848Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9519477Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9520083Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9520755Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9521381Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9521993Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9522596Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9523201Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9523805Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9524407Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9525022Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9525159Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:16.9525199Z Autotune Choices Stats:
2025-12-04T09:45:16.9525966Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:16.9526194Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9526360Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9526638Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9527267Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9527887Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9528514Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9529148Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9529798Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9530474Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9531094Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9531717Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9532345Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9532965Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9533092Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:16.9533184Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9533226Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9533263Z unimplemented []
2025-12-04T09:45:16.9533335Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9533435Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9534031Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9534081Z graph_break []
2025-12-04T09:45:16.9534154Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9534194Z Autotune Choices Stats:
2025-12-04T09:45:16.9534940Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:16.9535067Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9535183Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9535343Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9535964Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9536573Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9537181Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9537800Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9538426Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9539035Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9539657Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9540263Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9540900Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9541499Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9541653Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:16.9541693Z Autotune Choices Stats:
2025-12-04T09:45:16.9542455Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.9542700Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9542884Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9543166Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9543799Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9544443Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9545086Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9545737Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9546382Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9547027Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9547658Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9548288Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9548916Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9549537Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9549666Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:16.9549740Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9549782Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9549820Z unimplemented []
2025-12-04T09:45:16.9549880Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9549979Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9550621Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9550682Z graph_break []
2025-12-04T09:45:16.9550755Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9550795Z Autotune Choices Stats:
2025-12-04T09:45:16.9551559Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:16.9551700Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9551814Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9551974Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9552596Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9553201Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9553809Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9554422Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9555025Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9555649Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9556269Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9556877Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9557497Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9558098Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9558230Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:16.9558269Z Autotune Choices Stats:
2025-12-04T09:45:16.9559034Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.9559262Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9559427Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9559718Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9560362Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9561028Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9561647Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9562263Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9562893Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9563542Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9564190Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9564826Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9565451Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9566079Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9566208Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:16.9566282Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9566324Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9566362Z unimplemented []
2025-12-04T09:45:16.9566423Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9566524Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9567094Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9567132Z graph_break []
2025-12-04T09:45:16.9567223Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9567263Z Autotune Choices Stats:
2025-12-04T09:45:16.9568009Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:16.9568157Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9568283Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9568443Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9569056Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9569661Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9570282Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9570927Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9571547Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9572154Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9572789Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9573411Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9574018Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9574621Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9574751Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:16.9574793Z Autotune Choices Stats:
2025-12-04T09:45:16.9575553Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:16.9575773Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9575953Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9576247Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9576893Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9577528Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9578161Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9578788Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9579416Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9580075Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9580732Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9581390Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9582042Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9582674Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9582804Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:16.9582878Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9582921Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9582958Z unimplemented []
2025-12-04T09:45:16.9583019Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9583119Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9583691Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9583729Z graph_break []
2025-12-04T09:45:16.9583802Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9583842Z Autotune Choices Stats:
2025-12-04T09:45:16.9584604Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.9584742Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9584856Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9585016Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9585648Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9586261Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9586871Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9587497Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9588114Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9588742Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9589349Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9589986Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9590632Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9591234Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9591364Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:16.9591402Z Autotune Choices Stats:
2025-12-04T09:45:16.9592167Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:16.9592387Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9592551Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9592851Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9593482Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9594135Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9594773Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9595405Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9596039Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9596671Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9597302Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9597930Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9598581Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9599215Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9599344Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:16.9599419Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9599460Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9599498Z unimplemented []
2025-12-04T09:45:16.9599558Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9599660Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9600235Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9600276Z graph_break []
2025-12-04T09:45:16.9600350Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9600391Z Autotune Choices Stats:
2025-12-04T09:45:16.9601181Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:16.9601309Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9601447Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9601607Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9602244Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9602857Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9603461Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9604065Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9604685Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9605288Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9605908Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9606514Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9607138Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9607759Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9607886Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:16.9607928Z Autotune Choices Stats:
2025-12-04T09:45:16.9608688Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:16.9608906Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9609074Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9609351Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9609988Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9610650Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9611310Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9611942Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9612582Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9613207Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9613829Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9614465Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9615089Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9615738Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9615876Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:16.9615950Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9615993Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9616030Z unimplemented []
2025-12-04T09:45:16.9616091Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9616193Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9616771Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9616808Z graph_break []
2025-12-04T09:45:16.9616882Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9616921Z Autotune Choices Stats:
2025-12-04T09:45:16.9617659Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:16.9617787Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9617900Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9618060Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9618680Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9619306Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9619924Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9620573Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9621175Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9621780Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9622387Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9623022Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9623636Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9624261Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9624401Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:16.9624441Z Autotune Choices Stats:
2025-12-04T09:45:16.9625205Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.9625424Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9625590Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9625869Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9626494Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9627127Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9627744Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9628388Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9629024Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9629651Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9630273Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9630944Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9631602Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9632261Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9632403Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:16.9632479Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9632519Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9632558Z unimplemented []
2025-12-04T09:45:16.9632618Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9632721Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9633306Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9633343Z graph_break []
2025-12-04T09:45:16.9633417Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9633457Z Autotune Choices Stats:
2025-12-04T09:45:16.9634198Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:16.9634328Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9634444Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9634603Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9635218Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9635832Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9636458Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9637070Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9637692Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9638314Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9638921Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9639541Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9640164Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9640833Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9640975Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:16.9641016Z Autotune Choices Stats:
2025-12-04T09:45:16.9641777Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:16.9641997Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9642164Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9642447Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9643088Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9643712Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9644349Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9644990Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9645632Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9646254Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9646872Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9647502Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9648129Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9648761Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9648899Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:16.9648973Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9649016Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9649054Z unimplemented []
2025-12-04T09:45:16.9649116Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9649227Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9649816Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9649853Z graph_break []
2025-12-04T09:45:16.9649928Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9649967Z Autotune Choices Stats:
2025-12-04T09:45:16.9650769Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:16.9650898Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9651012Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9651173Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9651795Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9652413Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9653046Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9653675Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9654289Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9654908Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9655532Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9656138Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9656747Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9657366Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9657506Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:16.9657545Z Autotune Choices Stats:
2025-12-04T09:45:16.9658313Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:16.9658540Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9658706Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9658986Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9659619Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9660246Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9660893Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9661545Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9662203Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9662841Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9663467Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9664095Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9664733Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9665361Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9665490Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:16.9665574Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9665616Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9665652Z unimplemented []
2025-12-04T09:45:16.9665724Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9665823Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9666413Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9666460Z graph_break []
2025-12-04T09:45:16.9666534Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9666574Z Autotune Choices Stats:
2025-12-04T09:45:16.9667319Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:16.9667449Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9667568Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9667729Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9668347Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9668957Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9669572Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9670185Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9670848Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9671465Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9672075Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9672677Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9673280Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9673885Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9674026Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:16.9674067Z Autotune Choices Stats:
2025-12-04T09:45:16.9674842Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:16.9675084Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9675260Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9675540Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9676182Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9677514Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9678799Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9680085Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9681446Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9682773Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9684079Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9685356Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9686667Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9687965Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9688762Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:16.9688999Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9689156Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9689335Z unimplemented []
2025-12-04T09:45:16.9689529Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9689797Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9690584Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9691234Z graph_break []
2025-12-04T09:45:16.9691364Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9691519Z Autotune Choices Stats:
2025-12-04T09:45:16.9692358Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:16.9693274Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9693553Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9693866Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9694682Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9695942Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9697198Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9698458Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9699704Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9701014Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9702284Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9703529Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9704773Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9706018Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9706787Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:16.9706992Z Autotune Choices Stats:
2025-12-04T09:45:16.9707848Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:16.9708876Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9709299Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9709791Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9710797Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9712077Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9713363Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9714658Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9715952Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9717277Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9718595Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9719882Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9721193Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9722493Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9723294Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:16.9723532Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9723687Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9723797Z unimplemented []
2025-12-04T09:45:16.9723916Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9724113Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9724825Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9725465Z graph_break []
2025-12-04T09:45:16.9725617Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9725770Z Autotune Choices Stats:
2025-12-04T09:45:16.9726589Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:16.9727515Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9727795Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9728105Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9728919Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9730174Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9731457Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9732714Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9733981Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9735222Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9736499Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9737764Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9739016Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9740275Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9741077Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:16.9741281Z Autotune Choices Stats:
2025-12-04T09:45:16.9742111Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:16.9743151Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9743574Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9744065Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9745032Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9747241Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9748537Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9749833Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9751159Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9752461Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9753758Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9755090Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9756453Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9757745Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9758537Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:16.9758776Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9758932Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9759045Z unimplemented []
2025-12-04T09:45:16.9759165Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9759363Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9760076Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9760749Z graph_break []
2025-12-04T09:45:16.9760879Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9761031Z Autotune Choices Stats:
2025-12-04T09:45:16.9761843Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.9762773Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9763048Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9763381Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9764208Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9765505Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9766754Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9770974Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9772235Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9773479Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9774739Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9776059Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9777357Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9778601Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9779375Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:16.9779583Z Autotune Choices Stats:
2025-12-04T09:45:16.9780450Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.9781471Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9781905Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9782394Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9783351Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9784722Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9786057Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9787339Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9788643Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9789951Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9791289Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9792585Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9793906Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9795234Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9796038Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:16.9796291Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9796461Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9796585Z unimplemented []
2025-12-04T09:45:16.9796718Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9796927Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9797655Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9798305Z graph_break []
2025-12-04T09:45:16.9798433Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9798590Z Autotune Choices Stats:
2025-12-04T09:45:16.9799408Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:16.9800319Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9800649Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9800966Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9801833Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9803100Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9804378Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9805638Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9806898Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9808146Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9809408Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9810741Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9812015Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9813288Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9814076Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:16.9814293Z Autotune Choices Stats:
2025-12-04T09:45:16.9815129Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:16.9816149Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9816571Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9817051Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9818000Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9819305Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9820685Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9822020Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9823320Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9824627Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9825944Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9827241Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9828550Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9829850Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9830685Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:16.9830923Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9831084Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9831216Z unimplemented []
2025-12-04T09:45:16.9831337Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9831531Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9832246Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9832888Z graph_break []
2025-12-04T09:45:16.9833017Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9833169Z Autotune Choices Stats:
2025-12-04T09:45:16.9833975Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:16.9834880Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9835158Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9835466Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9836290Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9837594Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9838855Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9840119Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9841407Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9842642Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9843899Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9845165Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9846448Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9847717Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9848485Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:16.9848689Z Autotune Choices Stats:
2025-12-04T09:45:16.9849541Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:16.9850580Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9850997Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9851478Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9852421Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9853715Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9855048Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9856361Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9857665Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9858957Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9860253Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9861595Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9862890Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9864209Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9865006Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:16.9865245Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9865398Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9865507Z unimplemented []
2025-12-04T09:45:16.9865625Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9865820Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9866548Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9867186Z graph_break []
2025-12-04T09:45:16.9867315Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9867467Z Autotune Choices Stats:
2025-12-04T09:45:16.9868280Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:16.9869175Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9869449Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9869759Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9870622Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9871867Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9873154Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9874411Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9875686Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9876930Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9878180Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9879438Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9880722Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9881999Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9882787Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:16.9882989Z Autotune Choices Stats:
2025-12-04T09:45:16.9883838Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:16.9884841Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9885258Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9885735Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9886687Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9887982Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9889264Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9890617Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9891920Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9893235Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9894525Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9895928Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9897234Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9898519Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9899328Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:16.9899571Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9899725Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9899834Z unimplemented []
2025-12-04T09:45:16.9899965Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9900172Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9900917Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9901554Z graph_break []
2025-12-04T09:45:16.9901682Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9901832Z Autotune Choices Stats:
2025-12-04T09:45:16.9902676Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:16.9903582Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9903858Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9904167Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9904984Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9906260Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9907509Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9908802Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9910063Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9911388Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9912641Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9913897Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9915158Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9916420Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9917214Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:16.9917415Z Autotune Choices Stats:
2025-12-04T09:45:16.9918269Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:16.9919385Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9919820Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9920296Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9921293Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9922595Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9923880Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9925171Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9926488Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9927789Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9929094Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9930392Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9931744Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9933038Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9933835Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:16.9934073Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9934251Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9934359Z unimplemented []
2025-12-04T09:45:16.9934478Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9934676Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9935405Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:16.9936057Z graph_break []
2025-12-04T09:45:16.9936186Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9936337Z Autotune Choices Stats:
2025-12-04T09:45:16.9937164Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:16.9938061Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9938335Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9938644Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9939458Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9940760Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9942001Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9943245Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9944511Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9945778Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9947036Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9948294Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9949551Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9950845Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9951627Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:16.9951834Z Autotune Choices Stats:
2025-12-04T09:45:16.9952688Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:16.9953724Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9954144Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9954623Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9955589Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9956898Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9958192Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9959486Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9960818Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9962145Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9963472Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9964760Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9966052Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9967348Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9968142Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:16.9968384Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:16.9968541Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:16.9968650Z unimplemented []
2025-12-04T09:45:16.9968767Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:16.9968963Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:16.9969675Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:16.9970329Z graph_break []
2025-12-04T09:45:16.9970497Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:16.9970651Z Autotune Choices Stats:
2025-12-04T09:45:16.9971475Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:16.9972384Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9972674Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9972984Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9973810Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9975060Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9976305Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9977556Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9978798Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9980067Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:16.9981378Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9982630Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9983880Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9985132Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9985899Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:16.9986105Z Autotune Choices Stats:
2025-12-04T09:45:16.9986940Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:16.9987971Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:16.9988404Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:16.9988891Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:16.9989854Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9991255Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9992546Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9993835Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9995122Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9996422Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9997748Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:16.9999064Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0000361Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0001703Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0002495Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:17.0002737Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0002893Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0003001Z unimplemented []
2025-12-04T09:45:17.0003117Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0003315Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0004025Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0004668Z graph_break []
2025-12-04T09:45:17.0004797Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0004971Z Autotune Choices Stats:
2025-12-04T09:45:17.0005797Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:17.0006712Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0006987Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0007296Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0008119Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0009378Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0010699Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0011940Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0013189Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0014450Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0015714Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0016997Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0018245Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0019489Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0020257Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:17.0020505Z Autotune Choices Stats:
2025-12-04T09:45:17.0021330Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.0022362Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0022779Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0023274Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0024253Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0025579Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0026863Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0028152Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0029440Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0030764Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0032060Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0033412Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0034727Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0036029Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0036816Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:17.0037056Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0037212Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0037321Z unimplemented []
2025-12-04T09:45:17.0037440Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0037636Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0038347Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0038986Z graph_break []
2025-12-04T09:45:17.0039112Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0039266Z Autotune Choices Stats:
2025-12-04T09:45:17.0040081Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.0041039Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0041315Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0041642Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0042468Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0043746Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0044993Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0046240Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0047485Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0048748Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0050036Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0051324Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0052591Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0053826Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0054593Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:17.0054799Z Autotune Choices Stats:
2025-12-04T09:45:17.0055618Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:17.0056631Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0057049Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0057528Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0058507Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0059829Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0061195Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0062490Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0063770Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0065080Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0066381Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0067712Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0069033Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0070358Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0071182Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:17.0071420Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0071577Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0071686Z unimplemented []
2025-12-04T09:45:17.0071803Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0071998Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0072710Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0073345Z graph_break []
2025-12-04T09:45:17.0073474Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0073627Z Autotune Choices Stats:
2025-12-04T09:45:17.0074436Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.0075331Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0075605Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0075939Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0076771Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0078029Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0079310Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0080581Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0081826Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0083088Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0084343Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0085640Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0086884Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0088141Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0088911Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:17.0089115Z Autotune Choices Stats:
2025-12-04T09:45:17.0089934Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.0090982Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0091398Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0091871Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0092837Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0094165Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0095461Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0096763Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0098046Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0099349Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0100680Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0101964Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0103278Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0104596Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0105384Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:17.0105638Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0105794Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0105905Z unimplemented []
2025-12-04T09:45:17.0106024Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0106223Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0106935Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0107574Z graph_break []
2025-12-04T09:45:17.0107703Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0107854Z Autotune Choices Stats:
2025-12-04T09:45:17.0108661Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.0109564Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0109680Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0109845Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0110492Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0111121Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0111743Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0112381Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0112982Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0113590Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0114200Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0114822Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0115449Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0116059Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0116192Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:17.0116244Z Autotune Choices Stats:
2025-12-04T09:45:17.0117009Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.0117231Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0117399Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0117681Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0118327Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0118956Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0119606Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0120241Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0120935Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0121564Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0122195Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0122824Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0123468Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0124124Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0124268Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:17.0124345Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0124388Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0124426Z unimplemented []
2025-12-04T09:45:17.0124486Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0124589Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0125174Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0125213Z graph_break []
2025-12-04T09:45:17.0125286Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0125328Z Autotune Choices Stats:
2025-12-04T09:45:17.0126064Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1708", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.0126191Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0126307Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0126468Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0127084Z   triton_flex_attention_1708 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0127690Z   triton_flex_attention_1706 0.0107 ms 93.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0128317Z   triton_flex_attention_1709 0.0110 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0128938Z   triton_flex_attention_1704 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0129553Z   triton_flex_attention_1707 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0130159Z   triton_flex_attention_1705 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0130789Z   triton_flex_attention_1724 0.0146 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0131392Z   triton_flex_attention_1716 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0131999Z   triton_flex_attention_1722 0.0160 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0132641Z   triton_flex_attention_1702 0.0166 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0132782Z SingleProcess AUTOTUNE benchmarking takes 0.1988 seconds and 0.5275 seconds precompiling for 24 choices
2025-12-04T09:45:17.0132825Z Autotune Choices Stats:
2025-12-04T09:45:17.0133606Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1743", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.0133826Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0133994Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0134274Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0134902Z   triton_flex_attention_backward_1743 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0135526Z   triton_flex_attention_backward_1737 0.0208 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0136153Z   triton_flex_attention_backward_1734 0.0213 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0136800Z   triton_flex_attention_backward_1735 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0137440Z   triton_flex_attention_backward_1745 0.0232 ms 76.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0138080Z   triton_flex_attention_backward_1744 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0138707Z   triton_flex_attention_backward_1742 0.0249 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0139356Z   triton_flex_attention_backward_1747 0.0252 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0140005Z   triton_flex_attention_backward_1738 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0140667Z   triton_flex_attention_backward_1729 0.0264 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0140810Z SingleProcess AUTOTUNE benchmarking takes 0.2428 seconds and 0.7372 seconds precompiling for 22 choices
2025-12-04T09:45:17.0140897Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0140941Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0140993Z unimplemented []
2025-12-04T09:45:17.0141054Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0141153Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0141734Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0141771Z graph_break []
2025-12-04T09:45:17.0141860Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0141900Z Autotune Choices Stats:
2025-12-04T09:45:17.0142647Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1754", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.0142777Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0142893Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0143056Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0143688Z   triton_flex_attention_1754 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0144303Z   triton_flex_attention_1752 0.0110 ms 90.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0144915Z   triton_flex_attention_1755 0.0114 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0145542Z   triton_flex_attention_1753 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0146164Z   triton_flex_attention_1750 0.0125 ms 79.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0146773Z   triton_flex_attention_1751 0.0143 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0147387Z   triton_flex_attention_1770 0.0149 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0148003Z   triton_flex_attention_1762 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0148613Z   triton_flex_attention_1768 0.0163 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0149223Z   triton_flex_attention_1748 0.0170 ms 58.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0149365Z SingleProcess AUTOTUNE benchmarking takes 0.2060 seconds and 0.4503 seconds precompiling for 24 choices
2025-12-04T09:45:17.0149407Z Autotune Choices Stats:
2025-12-04T09:45:17.0150181Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1789", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.0150431Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0150625Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0150908Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0151544Z   triton_flex_attention_backward_1789 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0152177Z   triton_flex_attention_backward_1783 0.0209 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0152800Z   triton_flex_attention_backward_1780 0.0216 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0153428Z   triton_flex_attention_backward_1781 0.0217 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0154086Z   triton_flex_attention_backward_1791 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0154742Z   triton_flex_attention_backward_1790 0.0235 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0155371Z   triton_flex_attention_backward_1788 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0156005Z   triton_flex_attention_backward_1793 0.0255 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0156626Z   triton_flex_attention_backward_1775 0.0264 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0157263Z   triton_flex_attention_backward_1784 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0157393Z SingleProcess AUTOTUNE benchmarking takes 0.2498 seconds and 0.6949 seconds precompiling for 22 choices
2025-12-04T09:45:17.0157480Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0157522Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0157561Z unimplemented []
2025-12-04T09:45:17.0157622Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0157724Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0158313Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0158363Z graph_break []
2025-12-04T09:45:17.0158439Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0158482Z Autotune Choices Stats:
2025-12-04T09:45:17.0159231Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1800", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.0159360Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0159477Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0159639Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0160259Z   triton_flex_attention_1800 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0160902Z   triton_flex_attention_1798 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0161504Z   triton_flex_attention_1801 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0162140Z   triton_flex_attention_1796 0.0121 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0162766Z   triton_flex_attention_1799 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0163403Z   triton_flex_attention_1816 0.0145 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0164007Z   triton_flex_attention_1797 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0164616Z   triton_flex_attention_1808 0.0152 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0165221Z   triton_flex_attention_1814 0.0161 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0165827Z   triton_flex_attention_1806 0.0168 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0165955Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.5450 seconds precompiling for 24 choices
2025-12-04T09:45:17.0166008Z Autotune Choices Stats:
2025-12-04T09:45:17.0166782Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1835", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.0167013Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0167179Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0167468Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0168103Z   triton_flex_attention_backward_1835 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0168725Z   triton_flex_attention_backward_1829 0.0210 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0169356Z   triton_flex_attention_backward_1826 0.0212 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0169974Z   triton_flex_attention_backward_1827 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0170661Z   triton_flex_attention_backward_1837 0.0231 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0171320Z   triton_flex_attention_backward_1836 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0171977Z   triton_flex_attention_backward_1839 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0172603Z   triton_flex_attention_backward_1834 0.0252 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0173225Z   triton_flex_attention_backward_1830 0.0260 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0173851Z   triton_flex_attention_backward_1821 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0173980Z SingleProcess AUTOTUNE benchmarking takes 0.2508 seconds and 0.7770 seconds precompiling for 22 choices
2025-12-04T09:45:17.0174056Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0174101Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0174139Z unimplemented []
2025-12-04T09:45:17.0174199Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0174299Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0174878Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0174929Z graph_break []
2025-12-04T09:45:17.0175016Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0175066Z Autotune Choices Stats:
2025-12-04T09:45:17.0175805Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1846", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.0175933Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0176066Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0176228Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0176840Z   triton_flex_attention_1846 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0177453Z   triton_flex_attention_1844 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0178067Z   triton_flex_attention_1847 0.0114 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0178677Z   triton_flex_attention_1842 0.0122 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0179308Z   triton_flex_attention_1845 0.0124 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0179925Z   triton_flex_attention_1843 0.0144 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0180614Z   triton_flex_attention_1862 0.0146 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0181223Z   triton_flex_attention_1854 0.0154 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0181833Z   triton_flex_attention_1860 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0182440Z   triton_flex_attention_1840 0.0167 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0182571Z SingleProcess AUTOTUNE benchmarking takes 0.2278 seconds and 0.3492 seconds precompiling for 24 choices
2025-12-04T09:45:17.0182612Z Autotune Choices Stats:
2025-12-04T09:45:17.0183378Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1881", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.0183610Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0183787Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0184074Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0184723Z   triton_flex_attention_backward_1881 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0185355Z   triton_flex_attention_backward_1875 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0185983Z   triton_flex_attention_backward_1873 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0186624Z   triton_flex_attention_backward_1872 0.0216 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0187271Z   triton_flex_attention_backward_1882 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0187913Z   triton_flex_attention_backward_1883 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0188549Z   triton_flex_attention_backward_1880 0.0254 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0189199Z   triton_flex_attention_backward_1885 0.0254 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0189827Z   triton_flex_attention_backward_1876 0.0263 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0190485Z   triton_flex_attention_backward_1867 0.0267 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0190614Z SingleProcess AUTOTUNE benchmarking takes 0.2409 seconds and 0.8665 seconds precompiling for 22 choices
2025-12-04T09:45:17.0190689Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0190732Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0190769Z unimplemented []
2025-12-04T09:45:17.0190831Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0190930Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0191512Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 74), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 28), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 12), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0191551Z graph_break []
2025-12-04T09:45:17.0191653Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0191694Z Autotune Choices Stats:
2025-12-04T09:45:17.0192453Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1892", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.0192602Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0192717Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0192878Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0193511Z   triton_flex_attention_1892 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0194118Z   triton_flex_attention_1890 0.0109 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0194739Z   triton_flex_attention_1893 0.0114 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0195348Z   triton_flex_attention_1888 0.0122 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0195958Z   triton_flex_attention_1891 0.0123 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0196593Z   triton_flex_attention_1889 0.0144 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0197212Z   triton_flex_attention_1908 0.0146 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0197826Z   triton_flex_attention_1900 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0198431Z   triton_flex_attention_1906 0.0161 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0199033Z   triton_flex_attention_1886 0.0167 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0199163Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3466 seconds precompiling for 24 choices
2025-12-04T09:45:17.0199204Z Autotune Choices Stats:
2025-12-04T09:45:17.0199964Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1927", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.0200185Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0200364Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0200678Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0201336Z   triton_flex_attention_backward_1927 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0201990Z   triton_flex_attention_backward_1921 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0202616Z   triton_flex_attention_backward_1918 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0203249Z   triton_flex_attention_backward_1919 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0203903Z   triton_flex_attention_backward_1929 0.0231 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0204535Z   triton_flex_attention_backward_1928 0.0233 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0205192Z   triton_flex_attention_backward_1926 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0205834Z   triton_flex_attention_backward_1931 0.0254 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0206476Z   triton_flex_attention_backward_1922 0.0261 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0207103Z   triton_flex_attention_backward_1913 0.0263 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0207232Z SingleProcess AUTOTUNE benchmarking takes 0.2431 seconds and 0.7860 seconds precompiling for 22 choices
2025-12-04T09:45:17.0207329Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:17.0207377Z Traceback (most recent call last):
2025-12-04T09:45:17.0207538Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:17.0207578Z     self.assertTrue(
2025-12-04T09:45:17.0207688Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:17.0207739Z     raise self.failureException(msg)
2025-12-04T09:45:17.0207866Z AssertionError: False is not true : Log file /tmp/tmpqm289lwi/flex_attention_configs.json was not created
2025-12-04T09:45:17.0207870Z 
2025-12-04T09:45:17.0207948Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.0208116Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.0208119Z 
2025-12-04T09:45:17.0208212Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.0208287Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0208332Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0208369Z unimplemented []
2025-12-04T09:45:17.0208430Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0209008Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:17.0209123Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0209183Z graph_break []
2025-12-04T09:45:17.0209270Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0209765Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:17.0209815Z   current_size = base.storage().size()
2025-12-04T09:45:17.0209856Z Autotune Choices Stats:
2025-12-04T09:45:17.0210651Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.0210782Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0210898Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0211059Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0211677Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0212295Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0212900Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0213526Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0214141Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0214763Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0215375Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0215985Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0216592Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0217201Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0217345Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:17.0217390Z Autotune Choices Stats:
2025-12-04T09:45:17.0218171Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.0218404Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0218573Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0218871Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0219503Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0220130Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0220777Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0221405Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0222035Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0222690Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0223338Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0223971Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0224603Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0225227Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0225358Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:17.0225434Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0225480Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0225521Z unimplemented []
2025-12-04T09:45:17.0225580Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0225682Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0226273Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0226313Z graph_break []
2025-12-04T09:45:17.0226399Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0226450Z Autotune Choices Stats:
2025-12-04T09:45:17.0227188Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.0227319Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0227447Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0227611Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0228223Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0228831Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0229442Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0230064Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0230748Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0231365Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0231988Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0232592Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0233199Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0233801Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0233933Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:17.0233974Z Autotune Choices Stats:
2025-12-04T09:45:17.0234738Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.0234971Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0235152Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0235443Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0236085Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0236707Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0237339Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0237964Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0238597Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0239233Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0239874Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0240579Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0241208Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0241836Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0241967Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:17.0242043Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0242088Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0242126Z unimplemented []
2025-12-04T09:45:17.0242188Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0242291Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0242870Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0242909Z graph_break []
2025-12-04T09:45:17.0243000Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0243042Z Autotune Choices Stats:
2025-12-04T09:45:17.0243822Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.0243963Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0244078Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0244238Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0244869Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0245473Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0246095Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0246698Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0247303Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0247933Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0248549Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0249169Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0249780Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0250386Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0250544Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:17.0250587Z Autotune Choices Stats:
2025-12-04T09:45:17.0251354Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.0251573Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0251753Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0252035Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0252673Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0253324Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0253949Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0254586Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0255217Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0255848Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0256497Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0257128Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0257769Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0258393Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0258524Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:17.0258601Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0258644Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0258684Z unimplemented []
2025-12-04T09:45:17.0258745Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0258846Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0259429Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0259469Z graph_break []
2025-12-04T09:45:17.0259545Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0259589Z Autotune Choices Stats:
2025-12-04T09:45:17.0260331Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:17.0260495Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0260627Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0260804Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0261420Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0262045Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0262654Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0263265Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0263865Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0264472Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0265103Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0265721Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0266332Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0266938Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0267068Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:17.0267110Z Autotune Choices Stats:
2025-12-04T09:45:17.0267877Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.0268094Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0268264Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0268546Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0269200Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0269837Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0270586Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0271211Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0271846Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0272494Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0273125Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0273788Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0274425Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0275073Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0275204Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:17.0275279Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0275322Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0275359Z unimplemented []
2025-12-04T09:45:17.0275421Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0275523Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0276099Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0276137Z graph_break []
2025-12-04T09:45:17.0276213Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0276253Z Autotune Choices Stats:
2025-12-04T09:45:17.0277002Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:17.0277131Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0277244Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0277416Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0278042Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0278654Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0279274Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0279879Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0280537Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0281140Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0281743Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0282378Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0282995Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0283616Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0283748Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:17.0283788Z Autotune Choices Stats:
2025-12-04T09:45:17.0284569Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.0284790Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0284957Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0285240Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0285885Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0286531Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0287163Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0287803Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0288430Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0289064Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0289690Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0290314Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0291011Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0291639Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0291768Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:17.0291858Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0291901Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0291940Z unimplemented []
2025-12-04T09:45:17.0292000Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0292101Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0292679Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0292720Z graph_break []
2025-12-04T09:45:17.0292793Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0292836Z Autotune Choices Stats:
2025-12-04T09:45:17.0293571Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:17.0293698Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0293816Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0293977Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0294592Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0295224Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0295829Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0296443Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0297052Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0297672Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0298281Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0298891Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0299520Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0302933Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0303114Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:17.0303159Z Autotune Choices Stats:
2025-12-04T09:45:17.0303925Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.0304151Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0304325Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0304607Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0305249Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0305870Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0306525Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0307157Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0307801Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0308430Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0309052Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0309678Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0310300Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0310997Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0311141Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:17.0311223Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0311268Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0311306Z unimplemented []
2025-12-04T09:45:17.0311370Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0311475Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0312066Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0312105Z graph_break []
2025-12-04T09:45:17.0312181Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0312221Z Autotune Choices Stats:
2025-12-04T09:45:17.0312969Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:17.0313101Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0313217Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0313384Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0314002Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0314612Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0315242Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0315866Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0316466Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0317071Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0317684Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0318299Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0318906Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0319538Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0319685Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:17.0319728Z Autotune Choices Stats:
2025-12-04T09:45:17.0320527Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.0320748Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0320916Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0321198Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0321837Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0322484Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0323119Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0323768Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0324411Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0325051Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0325677Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0326308Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0326957Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0327581Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0327722Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:17.0327810Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0327865Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0327903Z unimplemented []
2025-12-04T09:45:17.0327965Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0328069Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0328654Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0328693Z graph_break []
2025-12-04T09:45:17.0328779Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0328822Z Autotune Choices Stats:
2025-12-04T09:45:17.0329559Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:17.0329688Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0329805Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0329970Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0330630Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0331242Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0331847Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0332485Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0333116Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0333720Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0334330Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0334932Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0335544Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0336149Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0336296Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:17.0336337Z Autotune Choices Stats:
2025-12-04T09:45:17.0337122Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.0337352Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0337532Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0337813Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0338446Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0339068Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0339689Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0340310Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0340997Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0341645Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0342267Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0342909Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0343534Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0344156Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0344286Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:17.0344385Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0344429Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0344466Z unimplemented []
2025-12-04T09:45:17.0344529Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0344630Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0345219Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0345266Z graph_break []
2025-12-04T09:45:17.0345341Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0345381Z Autotune Choices Stats:
2025-12-04T09:45:17.0346136Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.0346267Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0346382Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0346547Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0347164Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0347772Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0348374Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0348983Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0349600Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0350236Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0350881Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0351492Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0352093Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0352700Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0352831Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:17.0352887Z Autotune Choices Stats:
2025-12-04T09:45:17.0353660Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.0353891Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0354059Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0354355Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0354986Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0355617Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0356233Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0356856Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0357484Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0358139Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0358783Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0359410Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0360041Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0360721Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0360854Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:17.0360929Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0360974Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0361013Z unimplemented []
2025-12-04T09:45:17.0361075Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0361177Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0361757Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0361810Z graph_break []
2025-12-04T09:45:17.0361884Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0361938Z Autotune Choices Stats:
2025-12-04T09:45:17.0362698Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:17.0362827Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0362955Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0363120Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0363732Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0364339Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0364945Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0365544Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0366157Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0366772Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0367405Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0368014Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0368624Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0369241Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0369372Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:17.0369414Z Autotune Choices Stats:
2025-12-04T09:45:17.0370172Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.0370448Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0370637Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0370935Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0371592Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0372217Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0372839Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0373484Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0374112Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0374761Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0375406Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0376058Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0376684Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0377308Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0377436Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:17.0377512Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0377553Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0377593Z unimplemented []
2025-12-04T09:45:17.0377654Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0377756Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0378333Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0378371Z graph_break []
2025-12-04T09:45:17.0378446Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0378496Z Autotune Choices Stats:
2025-12-04T09:45:17.0379251Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:17.0379389Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0379503Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0379666Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0380293Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0380933Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0381539Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0382141Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0382747Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0383397Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0384017Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0384645Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0385253Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0385857Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0385986Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:17.0386027Z Autotune Choices Stats:
2025-12-04T09:45:17.0386795Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.0387014Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0387181Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0387471Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0388112Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0388762Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0389386Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0390011Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0390676Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0391303Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0391942Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0392586Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0393236Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0393866Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0393997Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:17.0394070Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0394115Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0394154Z unimplemented []
2025-12-04T09:45:17.0394215Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0394316Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0394892Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0394931Z graph_break []
2025-12-04T09:45:17.0395006Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0395048Z Autotune Choices Stats:
2025-12-04T09:45:17.0395788Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.0395927Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0396054Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0396226Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0396839Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0397457Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0398064Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0398671Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0399280Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0399884Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0400570Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0401187Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0401805Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0402407Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0402539Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:17.0402580Z Autotune Choices Stats:
2025-12-04T09:45:17.0403338Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.0403558Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0403728Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0404008Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0404657Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0405291Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0405936Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0406662Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0407294Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0407923Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0408546Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0409196Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0409837Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0410543Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0410673Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:17.0410750Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0410793Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0410832Z unimplemented []
2025-12-04T09:45:17.0410894Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0411003Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0411583Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0411623Z graph_break []
2025-12-04T09:45:17.0411698Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0411739Z Autotune Choices Stats:
2025-12-04T09:45:17.0412476Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.0412603Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0412718Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0412903Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0413547Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0414170Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0414792Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0415393Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0415993Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0416599Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0417208Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0417833Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0418450Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0419060Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0419191Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:17.0419232Z Autotune Choices Stats:
2025-12-04T09:45:17.0419993Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.0420211Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0420379Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0420699Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0421331Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0421980Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0422616Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0423280Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0423908Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0424534Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0425160Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0425804Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0426459Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0427091Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0427224Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:17.0427307Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0427351Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0427389Z unimplemented []
2025-12-04T09:45:17.0427451Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0427552Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0428127Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0428168Z graph_break []
2025-12-04T09:45:17.0428242Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0428284Z Autotune Choices Stats:
2025-12-04T09:45:17.0429028Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.0429163Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0429278Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0429443Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0430056Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0430731Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0431357Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0431981Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0432587Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0433187Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0433800Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0434403Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0435042Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0435652Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0435782Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:17.0435835Z Autotune Choices Stats:
2025-12-04T09:45:17.0436600Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.0436820Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0436987Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0437278Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0437927Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0438549Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0439194Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0439826Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0440504Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0441130Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0441751Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0442377Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0443007Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0443661Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0443802Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:17.0443876Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0443919Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0443960Z unimplemented []
2025-12-04T09:45:17.0444020Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0444121Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0444708Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0444747Z graph_break []
2025-12-04T09:45:17.0444821Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0444862Z Autotune Choices Stats:
2025-12-04T09:45:17.0445626Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:17.0445753Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0445869Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0446032Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0446652Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0447277Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0447918Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0448530Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0449148Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0449754Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0450360Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0451015Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0451624Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0452269Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0452409Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:17.0452453Z Autotune Choices Stats:
2025-12-04T09:45:17.0453223Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:17.0453441Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0453609Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0453888Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0454518Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0455143Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0455765Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0456412Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0457045Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0457682Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0458304Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0458934Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0459561Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0460182Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0460321Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:17.0460397Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0460483Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0460533Z unimplemented []
2025-12-04T09:45:17.0460594Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0460694Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0461268Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0461305Z graph_break []
2025-12-04T09:45:17.0461395Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0461434Z Autotune Choices Stats:
2025-12-04T09:45:17.0462180Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.0462308Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0462422Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0462585Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0463205Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0463809Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0464415Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0465050Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0465673Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0466275Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0466881Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0467489Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0468096Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0468695Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0468835Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:17.0468876Z Autotune Choices Stats:
2025-12-04T09:45:17.0469650Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.0469879Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0470055Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0470337Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0470995Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0471618Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0472243Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0472865Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0473524Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0474162Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0474800Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0475428Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0476053Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0476677Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0476807Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:17.0476881Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0476936Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0476976Z unimplemented []
2025-12-04T09:45:17.0477037Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0477138Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0477728Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0477777Z graph_break []
2025-12-04T09:45:17.0477850Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0477894Z Autotune Choices Stats:
2025-12-04T09:45:17.0478656Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.0478784Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0478899Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0479062Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0479675Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0480276Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0480907Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0481511Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0482144Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0482772Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0483380Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0484000Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0484618Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0485221Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0485350Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:17.0485391Z Autotune Choices Stats:
2025-12-04T09:45:17.0486175Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.0486403Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0486570Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0486851Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0487510Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0488144Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0488771Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0489404Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0490035Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0490718Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0491381Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0492010Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0492642Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0493266Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0493399Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:17.0493474Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0493519Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0493557Z unimplemented []
2025-12-04T09:45:17.0493618Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0493719Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0494300Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0494351Z graph_break []
2025-12-04T09:45:17.0494426Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0494466Z Autotune Choices Stats:
2025-12-04T09:45:17.0495216Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:17.0495363Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0495487Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0495650Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0496259Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0496864Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0497471Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0498091Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0498695Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0499316Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0499945Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0500593Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0501216Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0501822Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0501953Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:17.0501993Z Autotune Choices Stats:
2025-12-04T09:45:17.0502758Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:17.0502994Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0503174Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0503470Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0504125Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0504755Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0505411Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0506040Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0506676Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0507312Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0507963Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0508612Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0509240Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0509866Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0509996Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:17.0510071Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0510114Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0510154Z unimplemented []
2025-12-04T09:45:17.0510213Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0510313Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0510934Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0510973Z graph_break []
2025-12-04T09:45:17.0511047Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0511103Z Autotune Choices Stats:
2025-12-04T09:45:17.0511867Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.0512010Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0512128Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0512293Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0512918Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0513527Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0514136Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0514748Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0515357Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0515972Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0516606Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0517237Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0517842Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0518462Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0518593Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:17.0518634Z Autotune Choices Stats:
2025-12-04T09:45:17.0519402Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:17.0519624Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0519793Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0520097Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0520800Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0521463Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0522089Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0522714Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0523342Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0523975Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0524599Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0525254Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0525903Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0526532Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0526663Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:17.0526738Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0526782Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0526820Z unimplemented []
2025-12-04T09:45:17.0526882Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0526982Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0527557Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0527596Z graph_break []
2025-12-04T09:45:17.0527672Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0527713Z Autotune Choices Stats:
2025-12-04T09:45:17.0528457Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.0528602Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0528718Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0528892Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0529526Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0530144Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0530788Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0531401Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0532008Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0532618Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0533254Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0533872Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0534491Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0535094Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0535226Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:17.0535266Z Autotune Choices Stats:
2025-12-04T09:45:17.0536025Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.0536247Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0536413Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0536703Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0537345Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0537991Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0538645Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0539276Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0539908Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0540567Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0541189Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0541854Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0542492Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0543139Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0543270Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:17.0543347Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0543390Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0543429Z unimplemented []
2025-12-04T09:45:17.0543489Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0543589Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0544166Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0544205Z graph_break []
2025-12-04T09:45:17.0544277Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0544320Z Autotune Choices Stats:
2025-12-04T09:45:17.0545061Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.0545189Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0545303Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0545479Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0546100Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0546714Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0547334Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0547944Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0548547Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0549151Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0549757Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0550384Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0551086Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0551711Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0551840Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:17.0551881Z Autotune Choices Stats:
2025-12-04T09:45:17.0552646Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.0552869Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0553036Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0553323Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0553971Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0554629Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0555268Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0555907Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0556533Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0557177Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0557816Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0558444Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0559091Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0559731Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0559859Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:17.0559945Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0559988Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0560028Z unimplemented []
2025-12-04T09:45:17.0560089Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0560190Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0560817Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0560855Z graph_break []
2025-12-04T09:45:17.0560931Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0560973Z Autotune Choices Stats:
2025-12-04T09:45:17.0561711Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:17.0561840Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0561955Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0562118Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0562725Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0563354Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0563975Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0564594Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0565193Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0565793Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0566403Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0567024Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0567655Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0568270Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0568403Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:17.0568444Z Autotune Choices Stats:
2025-12-04T09:45:17.0569215Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.0569436Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0569602Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0569883Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0570573Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0571206Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0571861Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0572499Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0573143Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0573771Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0574395Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0575030Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0575652Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0576305Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0576444Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:17.0576518Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0576562Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0576599Z unimplemented []
2025-12-04T09:45:17.0576661Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0576762Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0577351Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0577392Z graph_break []
2025-12-04T09:45:17.0577464Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0577507Z Autotune Choices Stats:
2025-12-04T09:45:17.0578249Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.0578379Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0578495Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0578656Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0579267Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0579868Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0580532Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0581151Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0581767Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0582373Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0582980Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0583585Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0584191Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0584825Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0584964Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:17.0585005Z Autotune Choices Stats:
2025-12-04T09:45:17.0585775Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.0585996Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0586162Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0586449Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0587082Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0587708Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0588336Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0588986Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0589627Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0590265Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0590928Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0591560Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0592189Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0592817Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0592966Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:17.0593042Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0593096Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0593136Z unimplemented []
2025-12-04T09:45:17.0593209Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0593311Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0593888Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0593927Z graph_break []
2025-12-04T09:45:17.0594003Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0594058Z Autotune Choices Stats:
2025-12-04T09:45:17.0594805Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.0594937Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0595057Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0595220Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0595835Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0596440Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0597051Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0597678Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0598305Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0598910Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0599526Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0600131Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0600768Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0601387Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0601543Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:17.0601584Z Autotune Choices Stats:
2025-12-04T09:45:17.0602359Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.0602590Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0602772Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0603052Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0603686Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0604317Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0604961Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0605606Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0606258Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0606899Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0607534Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0608168Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0608810Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0609435Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0609567Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:17.0609640Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0609698Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0609736Z unimplemented []
2025-12-04T09:45:17.0609798Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0609899Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0610541Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0610593Z graph_break []
2025-12-04T09:45:17.0610666Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0610708Z Autotune Choices Stats:
2025-12-04T09:45:17.0611468Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.0611599Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0611714Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0611876Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0612494Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0613099Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0613713Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0614321Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0614966Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0615590Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0616202Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0616813Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0617417Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0618027Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0618159Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:17.0618217Z Autotune Choices Stats:
2025-12-04T09:45:17.0618991Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.0619219Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0619387Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0619666Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0620310Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0620969Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0621596Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0622234Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0622872Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0623542Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0624204Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0624832Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0625462Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0626110Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0626242Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:17.0626318Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0626362Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0626404Z unimplemented []
2025-12-04T09:45:17.0626465Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0626566Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0627143Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0627194Z graph_break []
2025-12-04T09:45:17.0627270Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0627310Z Autotune Choices Stats:
2025-12-04T09:45:17.0628075Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:17.0628213Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0628338Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0628503Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0629116Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0629727Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0630339Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0630983Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0631614Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0632252Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0632892Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0633507Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0634120Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0634728Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0634861Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:17.0634901Z Autotune Choices Stats:
2025-12-04T09:45:17.0635689Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:17.0635918Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0636098Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0636388Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0637035Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0637658Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0638291Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0638917Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0639546Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0640177Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0640883Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0641544Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0642173Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0642799Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0642930Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:17.0643004Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0643047Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0643087Z unimplemented []
2025-12-04T09:45:17.0643151Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0643250Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0643825Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0643863Z graph_break []
2025-12-04T09:45:17.0643936Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0643992Z Autotune Choices Stats:
2025-12-04T09:45:17.0644747Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.0644887Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0645000Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0645165Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0645796Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0646402Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0647014Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0647638Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0648260Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0648891Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0649513Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0650150Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0650783Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0651390Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0651522Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:17.0651564Z Autotune Choices Stats:
2025-12-04T09:45:17.0652326Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.0652548Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0652727Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0653007Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0653656Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0654313Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0654939Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0655573Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0656209Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0656845Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0657502Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0661087Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0661742Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0662367Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0662496Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:17.0662573Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0662615Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0662657Z unimplemented []
2025-12-04T09:45:17.0662718Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0662819Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0663391Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0663430Z graph_break []
2025-12-04T09:45:17.0663505Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0663546Z Autotune Choices Stats:
2025-12-04T09:45:17.0664287Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:17.0664426Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0664554Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0664729Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0665348Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0665965Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0666577Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0667190Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0667802Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0668409Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0669040Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0669655Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0670270Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0670924Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0671055Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:17.0671097Z Autotune Choices Stats:
2025-12-04T09:45:17.0671862Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:17.0672080Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0672249Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0672528Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0673185Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0673825Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0674473Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0675100Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0675730Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0676352Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0676977Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0677621Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0678255Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0679142Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0679324Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:17.0679415Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0679459Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0679511Z unimplemented []
2025-12-04T09:45:17.0679575Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0679696Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0680304Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0680343Z graph_break []
2025-12-04T09:45:17.0680442Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0680483Z Autotune Choices Stats:
2025-12-04T09:45:17.0681229Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.0681359Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0681472Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0681653Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0682280Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0682896Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0683516Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0684123Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0684728Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0685335Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0685944Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0686570Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0687187Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0687799Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0687934Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:17.0687975Z Autotune Choices Stats:
2025-12-04T09:45:17.0688743Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.0688963Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0689133Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0689418Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0690050Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0690737Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0691381Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0692019Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0692647Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0693294Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0693924Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0694551Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0695199Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0695833Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0695972Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:17.0696051Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0696095Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0696135Z unimplemented []
2025-12-04T09:45:17.0696196Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0696302Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0696879Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0696921Z graph_break []
2025-12-04T09:45:17.0696996Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0697039Z Autotune Choices Stats:
2025-12-04T09:45:17.0697788Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:17.0697917Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0698035Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0698201Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0698815Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0699446Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0700061Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0700748Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0701352Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0701958Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0702575Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0703185Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0703811Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0704430Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0704572Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:17.0704621Z Autotune Choices Stats:
2025-12-04T09:45:17.0705390Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:17.0705609Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0705779Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0706058Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0706702Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0707331Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0707983Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0708615Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0709255Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0709891Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0710562Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0711193Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0711815Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0712481Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0712624Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:17.0712701Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0712744Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0712782Z unimplemented []
2025-12-04T09:45:17.0712843Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0712943Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0713527Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0713566Z graph_break []
2025-12-04T09:45:17.0713645Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0713686Z Autotune Choices Stats:
2025-12-04T09:45:17.0714448Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.0714578Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0714694Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0714864Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0715502Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0716108Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0716743Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0717368Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0717974Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0718583Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0719191Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0719802Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0720465Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0721101Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0721249Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:17.0721291Z Autotune Choices Stats:
2025-12-04T09:45:17.0722065Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.0722286Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0722453Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0722735Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0723371Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0723999Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0724630Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0725286Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0725928Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0726568Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0727195Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0727843Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0728467Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0729089Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0729234Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:17.0729322Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0729375Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0729415Z unimplemented []
2025-12-04T09:45:17.0729476Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0729578Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0730156Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0730207Z graph_break []
2025-12-04T09:45:17.0730282Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0730327Z Autotune Choices Stats:
2025-12-04T09:45:17.0731114Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.0731244Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0731360Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0731523Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0732149Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0732772Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0733381Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0734018Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0734653Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0735258Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0735867Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0736478Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0737086Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0737686Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0737826Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:17.0737877Z Autotune Choices Stats:
2025-12-04T09:45:17.0738652Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.0738893Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0739067Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0739348Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0739987Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0740650Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0741276Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0741901Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0742564Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0743215Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0743843Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0744466Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0745092Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0745723Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0745854Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:17.0745939Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0745983Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0746021Z unimplemented []
2025-12-04T09:45:17.0746083Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0746184Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0746772Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0746821Z graph_break []
2025-12-04T09:45:17.0746897Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0746937Z Autotune Choices Stats:
2025-12-04T09:45:17.0747693Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:17.0747821Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0747936Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0748098Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0748713Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0749322Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0749945Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0750612Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0751227Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0751845Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0752458Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0753072Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0753694Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0754317Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0754451Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:17.0754504Z Autotune Choices Stats:
2025-12-04T09:45:17.0755272Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.0755507Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0755673Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0755974Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0756612Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0757258Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0757889Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0758532Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0759169Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0759817Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0760498Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0761128Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0761758Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0762386Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0762518Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:17.0762594Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0762637Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0762679Z unimplemented []
2025-12-04T09:45:17.0762740Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0762841Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0763430Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0763471Z graph_break []
2025-12-04T09:45:17.0763556Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0763613Z Autotune Choices Stats:
2025-12-04T09:45:17.0764351Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:17.0764494Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0764611Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0764774Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0765400Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0766010Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0766622Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0767230Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0767855Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0768470Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0769089Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0769705Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0770316Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0770962Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0771093Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:17.0771139Z Autotune Choices Stats:
2025-12-04T09:45:17.0771897Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.0772131Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0772322Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0772617Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0773260Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0773887Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0774511Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0775138Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0775768Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0776418Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0777053Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0777706Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0778335Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0778964Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0779095Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:17.0779171Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0779214Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0779252Z unimplemented []
2025-12-04T09:45:17.0779314Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0779414Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0779988Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0780039Z graph_break []
2025-12-04T09:45:17.0780114Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0780154Z Autotune Choices Stats:
2025-12-04T09:45:17.0780942Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.0781083Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0781198Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0781361Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0781987Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0782597Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0783223Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0783851Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0784464Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0785092Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0785713Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0786337Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0786946Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0787552Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0787685Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:17.0787727Z Autotune Choices Stats:
2025-12-04T09:45:17.0788492Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:17.0788714Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0788895Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0789186Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0789830Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0790507Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0791128Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0791757Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0792394Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0793027Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0793684Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0794331Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0794978Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0795602Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0795734Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:17.0795811Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0795853Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0795894Z unimplemented []
2025-12-04T09:45:17.0795955Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0796056Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0796631Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0796670Z graph_break []
2025-12-04T09:45:17.0796744Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0796786Z Autotune Choices Stats:
2025-12-04T09:45:17.0797541Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.0797684Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0797809Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0797979Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0798599Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0799223Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0799835Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0800476Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0801083Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0801689Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0802344Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0802963Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0803609Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0804218Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0804348Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:17.0804391Z Autotune Choices Stats:
2025-12-04T09:45:17.0805149Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.0805371Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0805538Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0805818Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0806473Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0807113Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0807752Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0808383Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0809011Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0809640Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0810261Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0810960Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0811598Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0812231Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0812361Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:17.0812438Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0812482Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0812519Z unimplemented []
2025-12-04T09:45:17.0812581Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0812683Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0813263Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0813301Z graph_break []
2025-12-04T09:45:17.0813379Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0813419Z Autotune Choices Stats:
2025-12-04T09:45:17.0814189Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.0816234Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0816381Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0816547Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0817191Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0817831Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0818441Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0819049Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0819655Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0820258Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0820912Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0821559Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0822183Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0822806Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0822941Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:17.0822982Z Autotune Choices Stats:
2025-12-04T09:45:17.0823747Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.0823968Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0824136Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0824423Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0825053Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0825701Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0826337Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0826977Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0827610Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0828243Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0828876Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0829496Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0830144Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0830836Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0830987Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:17.0831064Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0831109Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0831147Z unimplemented []
2025-12-04T09:45:17.0831212Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0831313Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0831890Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0831932Z graph_break []
2025-12-04T09:45:17.0832007Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0832049Z Autotune Choices Stats:
2025-12-04T09:45:17.0832800Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1708", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.0832933Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0833052Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0833215Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0833833Z   triton_flex_attention_1708 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0834473Z   triton_flex_attention_1706 0.0107 ms 93.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0835106Z   triton_flex_attention_1709 0.0110 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0835713Z   triton_flex_attention_1704 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0836320Z   triton_flex_attention_1707 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0836940Z   triton_flex_attention_1705 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0837552Z   triton_flex_attention_1724 0.0146 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0838154Z   triton_flex_attention_1716 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0838788Z   triton_flex_attention_1722 0.0160 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0839412Z   triton_flex_attention_1702 0.0166 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0839542Z SingleProcess AUTOTUNE benchmarking takes 0.1988 seconds and 0.5275 seconds precompiling for 24 choices
2025-12-04T09:45:17.0839585Z Autotune Choices Stats:
2025-12-04T09:45:17.0840344Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1743", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.0840605Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0840774Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0841052Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0841686Z   triton_flex_attention_backward_1743 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0842320Z   triton_flex_attention_backward_1737 0.0208 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0842981Z   triton_flex_attention_backward_1734 0.0213 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0843632Z   triton_flex_attention_backward_1735 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0844269Z   triton_flex_attention_backward_1745 0.0232 ms 76.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0844899Z   triton_flex_attention_backward_1744 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0845517Z   triton_flex_attention_backward_1742 0.0249 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0846145Z   triton_flex_attention_backward_1747 0.0252 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0846771Z   triton_flex_attention_backward_1738 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0847414Z   triton_flex_attention_backward_1729 0.0264 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0847554Z SingleProcess AUTOTUNE benchmarking takes 0.2428 seconds and 0.7372 seconds precompiling for 22 choices
2025-12-04T09:45:17.0847629Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0847672Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0847711Z unimplemented []
2025-12-04T09:45:17.0847772Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0847886Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0848462Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0848501Z graph_break []
2025-12-04T09:45:17.0848577Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0848616Z Autotune Choices Stats:
2025-12-04T09:45:17.0849388Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1754", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.0849518Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0849633Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0849798Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0850451Z   triton_flex_attention_1754 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0851076Z   triton_flex_attention_1752 0.0110 ms 90.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0851695Z   triton_flex_attention_1755 0.0114 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0852330Z   triton_flex_attention_1753 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0852935Z   triton_flex_attention_1750 0.0125 ms 79.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0853539Z   triton_flex_attention_1751 0.0143 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0854164Z   triton_flex_attention_1770 0.0149 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0854774Z   triton_flex_attention_1762 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0855382Z   triton_flex_attention_1768 0.0163 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0856010Z   triton_flex_attention_1748 0.0170 ms 58.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0856151Z SingleProcess AUTOTUNE benchmarking takes 0.2060 seconds and 0.4503 seconds precompiling for 24 choices
2025-12-04T09:45:17.0856191Z Autotune Choices Stats:
2025-12-04T09:45:17.0856957Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1789", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.0857178Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0857349Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0857641Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0858275Z   triton_flex_attention_backward_1789 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0858899Z   triton_flex_attention_backward_1783 0.0209 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0859526Z   triton_flex_attention_backward_1780 0.0216 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0860174Z   triton_flex_attention_backward_1781 0.0217 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0860867Z   triton_flex_attention_backward_1791 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0861497Z   triton_flex_attention_backward_1790 0.0235 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0862147Z   triton_flex_attention_backward_1788 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0862777Z   triton_flex_attention_backward_1793 0.0255 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0863401Z   triton_flex_attention_backward_1775 0.0264 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0864028Z   triton_flex_attention_backward_1784 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0864185Z SingleProcess AUTOTUNE benchmarking takes 0.2498 seconds and 0.6949 seconds precompiling for 22 choices
2025-12-04T09:45:17.0864272Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0864316Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0864354Z unimplemented []
2025-12-04T09:45:17.0864417Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0864517Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0865107Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0865146Z graph_break []
2025-12-04T09:45:17.0865220Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0865261Z Autotune Choices Stats:
2025-12-04T09:45:17.0866014Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1800", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.0866145Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0866263Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0866424Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0867061Z   triton_flex_attention_1800 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0867671Z   triton_flex_attention_1798 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0868308Z   triton_flex_attention_1801 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0868926Z   triton_flex_attention_1796 0.0121 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0869546Z   triton_flex_attention_1799 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0870155Z   triton_flex_attention_1816 0.0145 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0870825Z   triton_flex_attention_1797 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0871430Z   triton_flex_attention_1808 0.0152 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0872041Z   triton_flex_attention_1814 0.0161 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0872672Z   triton_flex_attention_1806 0.0168 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0872814Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.5450 seconds precompiling for 24 choices
2025-12-04T09:45:17.0872872Z Autotune Choices Stats:
2025-12-04T09:45:17.0873639Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1835", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.0873870Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0874038Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0874318Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0874955Z   triton_flex_attention_backward_1835 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0875588Z   triton_flex_attention_backward_1829 0.0210 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0876215Z   triton_flex_attention_backward_1826 0.0212 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0876843Z   triton_flex_attention_backward_1827 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0877500Z   triton_flex_attention_backward_1837 0.0231 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0878147Z   triton_flex_attention_backward_1836 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0878777Z   triton_flex_attention_backward_1839 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0879406Z   triton_flex_attention_backward_1834 0.0252 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0880030Z   triton_flex_attention_backward_1830 0.0260 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0880709Z   triton_flex_attention_backward_1821 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0880857Z SingleProcess AUTOTUNE benchmarking takes 0.2508 seconds and 0.7770 seconds precompiling for 22 choices
2025-12-04T09:45:17.0880933Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0880975Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0881014Z unimplemented []
2025-12-04T09:45:17.0881075Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0881194Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0881782Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0881820Z graph_break []
2025-12-04T09:45:17.0881896Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0881936Z Autotune Choices Stats:
2025-12-04T09:45:17.0882704Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1846", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.0882833Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0882948Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0883110Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0883721Z   triton_flex_attention_1846 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0884334Z   triton_flex_attention_1844 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0884942Z   triton_flex_attention_1847 0.0114 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0885577Z   triton_flex_attention_1842 0.0122 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0886200Z   triton_flex_attention_1845 0.0124 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0886821Z   triton_flex_attention_1843 0.0144 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0887431Z   triton_flex_attention_1862 0.0146 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0888040Z   triton_flex_attention_1854 0.0154 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0888653Z   triton_flex_attention_1860 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0889262Z   triton_flex_attention_1840 0.0167 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0889408Z SingleProcess AUTOTUNE benchmarking takes 0.2278 seconds and 0.3492 seconds precompiling for 24 choices
2025-12-04T09:45:17.0889448Z Autotune Choices Stats:
2025-12-04T09:45:17.0890228Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1881", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.0890483Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0890652Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0890953Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0891583Z   triton_flex_attention_backward_1881 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0892211Z   triton_flex_attention_backward_1875 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0892833Z   triton_flex_attention_backward_1873 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0893478Z   triton_flex_attention_backward_1872 0.0216 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0894144Z   triton_flex_attention_backward_1882 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0894787Z   triton_flex_attention_backward_1883 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0895424Z   triton_flex_attention_backward_1880 0.0254 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0896056Z   triton_flex_attention_backward_1885 0.0254 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0896684Z   triton_flex_attention_backward_1876 0.0263 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0897312Z   triton_flex_attention_backward_1867 0.0267 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0897444Z SingleProcess AUTOTUNE benchmarking takes 0.2409 seconds and 0.8665 seconds precompiling for 22 choices
2025-12-04T09:45:17.0897518Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0897562Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0897599Z unimplemented []
2025-12-04T09:45:17.0897663Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0897773Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0898363Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 74), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 28), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 12), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0898411Z graph_break []
2025-12-04T09:45:17.0898485Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0898526Z Autotune Choices Stats:
2025-12-04T09:45:17.0899282Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1892", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.0899413Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0899528Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0899689Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0900310Z   triton_flex_attention_1892 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0900955Z   triton_flex_attention_1890 0.0109 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0901568Z   triton_flex_attention_1893 0.0114 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0902178Z   triton_flex_attention_1888 0.0122 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0902830Z   triton_flex_attention_1891 0.0123 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0903446Z   triton_flex_attention_1889 0.0144 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0904077Z   triton_flex_attention_1908 0.0146 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0904687Z   triton_flex_attention_1900 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0905292Z   triton_flex_attention_1906 0.0161 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0905895Z   triton_flex_attention_1886 0.0167 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0906026Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3466 seconds precompiling for 24 choices
2025-12-04T09:45:17.0906069Z Autotune Choices Stats:
2025-12-04T09:45:17.0906828Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1927", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.0907068Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0907245Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0907524Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0908171Z   triton_flex_attention_backward_1927 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0908798Z   triton_flex_attention_backward_1921 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0909423Z   triton_flex_attention_backward_1918 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0910048Z   triton_flex_attention_backward_1919 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0910708Z   triton_flex_attention_backward_1929 0.0231 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0911372Z   triton_flex_attention_backward_1928 0.0233 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0912005Z   triton_flex_attention_backward_1926 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0912649Z   triton_flex_attention_backward_1931 0.0254 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0913277Z   triton_flex_attention_backward_1922 0.0261 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0913898Z   triton_flex_attention_backward_1913 0.0263 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0914027Z SingleProcess AUTOTUNE benchmarking takes 0.2431 seconds and 0.7860 seconds precompiling for 22 choices
2025-12-04T09:45:17.0914102Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0914143Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0914182Z unimplemented []
2025-12-04T09:45:17.0914243Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0914346Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0914921Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0914972Z graph_break []
2025-12-04T09:45:17.0915047Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0915087Z Autotune Choices Stats:
2025-12-04T09:45:17.0915842Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1938", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.0915978Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0916094Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0916260Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0916881Z   triton_flex_attention_1938 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0917491Z   triton_flex_attention_1936 0.0109 ms 93.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0918100Z   triton_flex_attention_1939 0.0116 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0918703Z   triton_flex_attention_1934 0.0122 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0919317Z   triton_flex_attention_1937 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0919949Z   triton_flex_attention_1935 0.0144 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0920627Z   triton_flex_attention_1954 0.0148 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0921247Z   triton_flex_attention_1946 0.0154 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0921856Z   triton_flex_attention_1952 0.0164 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0922463Z   triton_flex_attention_1944 0.0170 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0922595Z SingleProcess AUTOTUNE benchmarking takes 0.2077 seconds and 0.3245 seconds precompiling for 24 choices
2025-12-04T09:45:17.0922635Z Autotune Choices Stats:
2025-12-04T09:45:17.0923399Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1973", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.0923631Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0923798Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0924090Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0924737Z   triton_flex_attention_backward_1973 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0925374Z   triton_flex_attention_backward_1967 0.0211 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0925999Z   triton_flex_attention_backward_1965 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0926641Z   triton_flex_attention_backward_1964 0.0217 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0927274Z   triton_flex_attention_backward_1975 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0927905Z   triton_flex_attention_backward_1974 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0928547Z   triton_flex_attention_backward_1972 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0929189Z   triton_flex_attention_backward_1977 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0929830Z   triton_flex_attention_backward_1968 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0930493Z   triton_flex_attention_backward_1959 0.0266 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0930624Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.8096 seconds precompiling for 22 choices
2025-12-04T09:45:17.0930716Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:17.0930766Z Traceback (most recent call last):
2025-12-04T09:45:17.0930921Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:17.0930964Z     self.assertTrue(
2025-12-04T09:45:17.0931075Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:17.0931125Z     raise self.failureException(msg)
2025-12-04T09:45:17.0931254Z AssertionError: False is not true : Log file /tmp/tmp6v5mhi1a/flex_attention_configs.json was not created
2025-12-04T09:45:17.0931258Z 
2025-12-04T09:45:17.0931337Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.0931505Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.0931510Z 
2025-12-04T09:45:17.0931600Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.0931677Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0931719Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0931779Z unimplemented []
2025-12-04T09:45:17.0931840Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0932433Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:17.0932545Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0932583Z graph_break []
2025-12-04T09:45:17.0932657Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0933151Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:17.0933200Z   current_size = base.storage().size()
2025-12-04T09:45:17.0933242Z Autotune Choices Stats:
2025-12-04T09:45:17.0934008Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.0934138Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0934254Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0934417Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0935026Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0935634Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0936255Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0936879Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0937485Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0938096Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0938706Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0939312Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0939913Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0940549Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0940697Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:17.0940739Z Autotune Choices Stats:
2025-12-04T09:45:17.0941517Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.0941751Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0941929Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0942208Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0942843Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0943487Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0944126Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0944747Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0945395Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0946029Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0946654Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0947278Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0947914Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0948532Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0948663Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:17.0948737Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0948780Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0948817Z unimplemented []
2025-12-04T09:45:17.0948879Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0948989Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0949579Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0949631Z graph_break []
2025-12-04T09:45:17.0949706Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0949746Z Autotune Choices Stats:
2025-12-04T09:45:17.0950542Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.0950672Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0950787Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0950951Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0951560Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0952162Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0952766Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0953363Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0953987Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0954607Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0955234Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0955839Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0956460Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0957080Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0957213Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:17.0957253Z Autotune Choices Stats:
2025-12-04T09:45:17.0958015Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.0958257Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0958431Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0958714Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0959358Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0959981Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0960641Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0961271Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0961905Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0962554Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0963186Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0963826Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0964453Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0965075Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0965207Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:17.0965282Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0965323Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0965362Z unimplemented []
2025-12-04T09:45:17.0965423Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0965525Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0966102Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0966151Z graph_break []
2025-12-04T09:45:17.0966225Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0966266Z Autotune Choices Stats:
2025-12-04T09:45:17.0967016Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.0967153Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0967269Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0967429Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0968047Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0968657Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0969265Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0969874Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0970498Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0971132Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0971744Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0972364Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0972970Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0973572Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0973704Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:17.0973746Z Autotune Choices Stats:
2025-12-04T09:45:17.0974514Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.0974741Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0974907Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0975193Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0975838Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0976477Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0977101Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0977729Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0978377Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0979024Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0979671Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0980303Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0980963Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0981589Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0981725Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:17.0981801Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0981844Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0981881Z unimplemented []
2025-12-04T09:45:17.0981943Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0982044Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0982634Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.0982672Z graph_break []
2025-12-04T09:45:17.0982747Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0982787Z Autotune Choices Stats:
2025-12-04T09:45:17.0983527Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:17.0983687Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0983813Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0983978Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0984599Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0985199Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0985809Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0986418Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0987021Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0987631Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0988257Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0988874Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0989504Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0990113Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0990244Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:17.0990284Z Autotune Choices Stats:
2025-12-04T09:45:17.0991053Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.0991273Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.0991440Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.0991734Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.0992381Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0993015Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0993645Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0994271Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0994917Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0995551Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0996172Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0996828Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.0997462Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0998101Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.0998231Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:17.0998306Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.0998348Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.0998387Z unimplemented []
2025-12-04T09:45:17.0998447Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.0998550Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.0999125Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.0999164Z graph_break []
2025-12-04T09:45:17.0999238Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.0999279Z Autotune Choices Stats:
2025-12-04T09:45:17.1000025Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:17.1000164Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1000280Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1000478Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1001116Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1001745Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1002353Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1002962Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1003575Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1004182Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1004784Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1005411Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1006035Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1006638Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1006769Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:17.1006810Z Autotune Choices Stats:
2025-12-04T09:45:17.1007576Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.1007799Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1007966Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1008244Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1008879Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1009524Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1010162Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1010844Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1011472Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1012101Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1012723Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1013348Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1013993Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1014646Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1014775Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:17.1014851Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1014893Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1014931Z unimplemented []
2025-12-04T09:45:17.1014993Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1015092Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1015667Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1015705Z graph_break []
2025-12-04T09:45:17.1015779Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1015819Z Autotune Choices Stats:
2025-12-04T09:45:17.1016572Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:17.1016703Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1016818Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1016979Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1017605Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1018221Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1018851Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1019457Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1020061Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1020696Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1021316Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1021931Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1022571Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1023195Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1023327Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:17.1023368Z Autotune Choices Stats:
2025-12-04T09:45:17.1024128Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.1024349Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1024515Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1024793Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1025447Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1026086Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1026732Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1027376Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1028005Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1028634Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1029253Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1029882Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1030542Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1031208Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1031348Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:17.1031423Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1031467Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1031504Z unimplemented []
2025-12-04T09:45:17.1031566Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1031681Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1032258Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1032298Z graph_break []
2025-12-04T09:45:17.1032371Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1032413Z Autotune Choices Stats:
2025-12-04T09:45:17.1033156Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:17.1033286Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1033401Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1033561Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1034177Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1034799Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1035436Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1036064Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1036671Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1037281Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1037888Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1038494Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1039102Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1039734Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1039873Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:17.1039914Z Autotune Choices Stats:
2025-12-04T09:45:17.1040735Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.1040955Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1041125Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1041406Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1042058Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1042688Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1043334Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1043987Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1044641Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1045265Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1045884Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1046511Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1047140Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1047763Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1047920Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:17.1048007Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1048048Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1048087Z unimplemented []
2025-12-04T09:45:17.1048148Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1048248Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1048838Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1048877Z graph_break []
2025-12-04T09:45:17.1048952Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1048992Z Autotune Choices Stats:
2025-12-04T09:45:17.1049730Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:17.1049859Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1049974Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1050135Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1050788Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1051392Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1052032Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1052648Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1053263Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1053864Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1054470Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1055076Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1055682Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1056299Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1056437Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:17.1056487Z Autotune Choices Stats:
2025-12-04T09:45:17.1057247Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.1057476Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1057644Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1057926Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1058568Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1059207Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1059836Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1060500Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1061167Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1061826Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1062452Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1063081Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1063706Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1064326Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1064468Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:17.1064541Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1064585Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1064622Z unimplemented []
2025-12-04T09:45:17.1064683Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1064784Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1065366Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1065416Z graph_break []
2025-12-04T09:45:17.1065490Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1065531Z Autotune Choices Stats:
2025-12-04T09:45:17.1066287Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.1066416Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1066532Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1066693Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1067308Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1067910Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1068527Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1069154Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1069768Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1070379Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1071027Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1071637Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1072238Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1072845Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1073004Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:17.1073045Z Autotune Choices Stats:
2025-12-04T09:45:17.1073819Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.1074049Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1074217Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1074512Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1075145Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1075770Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1076393Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1077018Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1077670Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1078316Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1078949Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1079577Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1080204Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1080870Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1080999Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:17.1081075Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1081117Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1081156Z unimplemented []
2025-12-04T09:45:17.1081217Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1081318Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1081918Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1081981Z graph_break []
2025-12-04T09:45:17.1082067Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1082106Z Autotune Choices Stats:
2025-12-04T09:45:17.1082843Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:17.1082982Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1083097Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1083259Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1083879Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1084488Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1085101Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1085703Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1086331Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1086947Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1087567Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1088172Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1088777Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1089398Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1089532Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:17.1089572Z Autotune Choices Stats:
2025-12-04T09:45:17.1090331Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.1090622Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1090810Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1092681Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1093342Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1093971Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1095402Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1096046Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1096680Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1097326Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1097946Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1098618Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1099242Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1099881Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1100013Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:17.1100088Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1100132Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1100170Z unimplemented []
2025-12-04T09:45:17.1100233Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1100334Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1100954Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1100994Z graph_break []
2025-12-04T09:45:17.1101067Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1101108Z Autotune Choices Stats:
2025-12-04T09:45:17.1101890Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:17.1102044Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1102161Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1102325Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1102955Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1103562Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1104186Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1104795Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1105392Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1106011Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1106618Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1107242Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1107845Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1108458Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1108589Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:17.1108630Z Autotune Choices Stats:
2025-12-04T09:45:17.1109393Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.1109612Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1109781Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1110076Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1110741Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1111400Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1112023Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1112657Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1113286Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1113914Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1114547Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1115177Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1115820Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1116446Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1116588Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:17.1116663Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1116705Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1116745Z unimplemented []
2025-12-04T09:45:17.1116805Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1116906Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1117488Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1117528Z graph_break []
2025-12-04T09:45:17.1117603Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1117643Z Autotune Choices Stats:
2025-12-04T09:45:17.1118398Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.1118536Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1118651Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1118812Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1119421Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1120048Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1120700Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1121319Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1121927Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1122534Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1123157Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1123759Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1124394Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1124997Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1125138Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:17.1125179Z Autotune Choices Stats:
2025-12-04T09:45:17.1125940Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.1126159Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1126328Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1126605Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1127251Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1127880Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1128530Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1129154Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1129791Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1130449Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1131072Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1131720Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1132347Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1132999Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1133129Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:17.1133202Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1133246Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1133284Z unimplemented []
2025-12-04T09:45:17.1133346Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1133466Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1134040Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1134079Z graph_break []
2025-12-04T09:45:17.1134153Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1134194Z Autotune Choices Stats:
2025-12-04T09:45:17.1134945Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.1135075Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1135191Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1135352Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1135984Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1136610Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1137216Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1137818Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1138431Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1139033Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1139637Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1140255Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1140894Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1141529Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1141659Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:17.1141701Z Autotune Choices Stats:
2025-12-04T09:45:17.1142455Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.1142687Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1142855Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1143136Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1143766Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1144414Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1145038Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1145682Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1146311Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1146948Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1147569Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1148196Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1148836Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1149464Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1149614Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:17.1149690Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1149732Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1149772Z unimplemented []
2025-12-04T09:45:17.1149832Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1149932Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1150539Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1150594Z graph_break []
2025-12-04T09:45:17.1150671Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1150710Z Autotune Choices Stats:
2025-12-04T09:45:17.1151455Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.1151585Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1151701Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1151862Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1152477Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1153100Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1153733Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1154336Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1154943Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1155561Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1156166Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1156768Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1157401Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1158030Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1158162Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:17.1158201Z Autotune Choices Stats:
2025-12-04T09:45:17.1158961Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.1159207Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1159373Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1159657Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1160291Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1160949Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1161591Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1162215Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1162870Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1163500Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1164137Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1164771Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1165396Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1166032Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1166175Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:17.1166249Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1166292Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1166329Z unimplemented []
2025-12-04T09:45:17.1166390Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1166500Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1167079Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1167119Z graph_break []
2025-12-04T09:45:17.1167192Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1167233Z Autotune Choices Stats:
2025-12-04T09:45:17.1167977Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:17.1168124Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1168237Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1168400Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1169018Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1169622Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1170237Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1170930Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1171530Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1172134Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1172762Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1173365Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1173968Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1174581Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1174726Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:17.1174766Z Autotune Choices Stats:
2025-12-04T09:45:17.1175546Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:17.1175767Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1175933Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1176228Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1176859Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1177487Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1178118Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1178758Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1179404Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1180034Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1180686Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1181329Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1181959Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1182584Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1182726Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:17.1182803Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1182844Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1182883Z unimplemented []
2025-12-04T09:45:17.1182944Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1183044Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1183649Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1183690Z graph_break []
2025-12-04T09:45:17.1183765Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1183806Z Autotune Choices Stats:
2025-12-04T09:45:17.1184546Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.1184684Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1184799Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1184960Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1185582Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1186190Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1186816Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1187415Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1188036Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1188641Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1189253Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1189864Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1190524Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1191130Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1191281Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:17.1191323Z Autotune Choices Stats:
2025-12-04T09:45:17.1192084Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.1192333Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1192500Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1192779Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1193413Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1194051Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1194679Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1195303Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1195946Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1196595Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1197219Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1197850Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1198488Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1199114Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1199245Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:17.1199320Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1199361Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1199400Z unimplemented []
2025-12-04T09:45:17.1199460Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1199562Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1200151Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1200201Z graph_break []
2025-12-04T09:45:17.1200274Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1200316Z Autotune Choices Stats:
2025-12-04T09:45:17.1201113Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.1201242Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1201358Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1201519Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1202147Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1202760Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1203381Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1204001Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1204604Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1205231Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1205838Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1206452Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1207068Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1207667Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1207796Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:17.1207838Z Autotune Choices Stats:
2025-12-04T09:45:17.1208609Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.1208828Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1209007Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1209297Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1209923Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1210585Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1211234Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1211872Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1212506Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1213155Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1213800Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1214428Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1215054Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1215690Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1215821Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:17.1215894Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1215937Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1215974Z unimplemented []
2025-12-04T09:45:17.1216036Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1216136Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1216710Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1216749Z graph_break []
2025-12-04T09:45:17.1216834Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1216874Z Autotune Choices Stats:
2025-12-04T09:45:17.1217624Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:17.1217774Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1217890Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1218055Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1218667Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1219283Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1219891Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1220530Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1221156Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1221759Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1222395Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1222998Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1223624Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1224228Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1224360Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:17.1224400Z Autotune Choices Stats:
2025-12-04T09:45:17.1225165Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:17.1225397Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1225563Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1225839Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1226500Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1227127Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1227755Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1228392Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1229024Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1229662Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1230278Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1230960Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1231591Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1232226Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1232357Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:17.1232431Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1232472Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1232511Z unimplemented []
2025-12-04T09:45:17.1232571Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1232673Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1233245Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1233284Z graph_break []
2025-12-04T09:45:17.1233357Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1233398Z Autotune Choices Stats:
2025-12-04T09:45:17.1234152Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.1234280Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1234410Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1234571Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1235195Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1235804Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1236421Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1237026Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1237645Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1238263Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1238867Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1239502Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1240107Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1240756Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1240885Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:17.1240927Z Autotune Choices Stats:
2025-12-04T09:45:17.1241685Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:17.1241907Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1242075Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1242374Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1243009Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1243659Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1244285Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1244922Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1245548Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1246170Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1246799Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1247431Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1248071Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1248696Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1248835Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:17.1248912Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1248954Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1248993Z unimplemented []
2025-12-04T09:45:17.1249054Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1249154Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1249733Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1249772Z graph_break []
2025-12-04T09:45:17.1249848Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1249887Z Autotune Choices Stats:
2025-12-04T09:45:17.1250690Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.1250822Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1250952Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1251115Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1251730Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1252358Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1252964Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1253582Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1254184Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1254786Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1255397Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1256003Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1256630Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1257231Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1257372Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:17.1257412Z Autotune Choices Stats:
2025-12-04T09:45:17.1258171Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.1258394Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1258560Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1258840Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1259489Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1260114Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1260818Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1261439Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1262083Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1262715Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1263334Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1263988Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1264616Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1265280Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1265411Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:17.1265484Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1265527Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1265565Z unimplemented []
2025-12-04T09:45:17.1265627Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1265740Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1266321Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1266359Z graph_break []
2025-12-04T09:45:17.1266432Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1266473Z Autotune Choices Stats:
2025-12-04T09:45:17.1267222Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.1267351Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1267468Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1267627Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1268257Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1268873Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1269507Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1270110Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1270840Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1271453Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1272058Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1272686Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1273290Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1273921Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1274050Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:17.1274092Z Autotune Choices Stats:
2025-12-04T09:45:17.1274856Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.1275093Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1275261Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1275542Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1276174Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1276830Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1277453Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1278096Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1278728Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1279365Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1279987Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1280671Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1281316Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1281944Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1282099Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:17.1282176Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1282218Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1282256Z unimplemented []
2025-12-04T09:45:17.1282316Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1282417Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1282999Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1283052Z graph_break []
2025-12-04T09:45:17.1283129Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1283169Z Autotune Choices Stats:
2025-12-04T09:45:17.1283903Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:17.1284032Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1284149Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1284314Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1284934Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1285567Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1286193Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1286796Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1287404Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1288020Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1288643Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1289244Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1289859Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1290492Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1290650Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:17.1290690Z Autotune Choices Stats:
2025-12-04T09:45:17.1291456Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.1291689Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1291859Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1292142Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1292778Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1293405Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1294039Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1294664Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1295313Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1295943Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1296582Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1297206Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1297832Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1298474Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1298612Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:17.1298685Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1298729Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1298766Z unimplemented []
2025-12-04T09:45:17.1298828Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1298929Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1299518Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1299558Z graph_break []
2025-12-04T09:45:17.1299631Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1299672Z Autotune Choices Stats:
2025-12-04T09:45:17.1300455Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.1300601Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1300716Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1300880Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1301503Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1302130Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1302772Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1303417Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1304021Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1304626Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1305248Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1305856Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1306470Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1307108Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1307252Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:17.1307294Z Autotune Choices Stats:
2025-12-04T09:45:17.1308068Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.1308288Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1308454Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1308745Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1309377Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1310011Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1310651Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1311295Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1311953Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1312588Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1313213Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1313852Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1314486Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1315135Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1315276Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:17.1315351Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1315393Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1315432Z unimplemented []
2025-12-04T09:45:17.1315492Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1315593Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1316199Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1316239Z graph_break []
2025-12-04T09:45:17.1316315Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1316355Z Autotune Choices Stats:
2025-12-04T09:45:17.1317101Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.1317239Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1317356Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1317518Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1318133Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1318746Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1319355Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1319973Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1320669Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1321275Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1321900Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1322515Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1323126Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1323733Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1323884Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:17.1327379Z Autotune Choices Stats:
2025-12-04T09:45:17.1328156Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.1328432Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1328607Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1328886Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1329526Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1330161Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1330825Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1331447Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1332096Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1332740Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1333363Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1333990Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1334629Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1335249Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1335384Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:17.1335465Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1335511Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1335550Z unimplemented []
2025-12-04T09:45:17.1335613Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1335717Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1336310Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1336359Z graph_break []
2025-12-04T09:45:17.1336437Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1336479Z Autotune Choices Stats:
2025-12-04T09:45:17.1337234Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.1337366Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1337484Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1337648Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1338273Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1338880Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1339485Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1340102Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1340734Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1341354Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1341955Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1342561Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1343182Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1343785Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1343915Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:17.1343955Z Autotune Choices Stats:
2025-12-04T09:45:17.1344722Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.1344941Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1345122Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1345408Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1346031Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1346657Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1347290Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1347914Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1348539Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1349178Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1349817Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1350492Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1351118Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1351758Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1351891Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:17.1351967Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1352010Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1352049Z unimplemented []
2025-12-04T09:45:17.1352110Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1352212Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1352788Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1352840Z graph_break []
2025-12-04T09:45:17.1352913Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1352955Z Autotune Choices Stats:
2025-12-04T09:45:17.1353689Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:17.1353842Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1353960Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1354122Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1354742Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1355366Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1355978Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1356581Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1357189Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1357797Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1358430Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1359034Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1359650Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1360258Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1360390Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:17.1360459Z Autotune Choices Stats:
2025-12-04T09:45:17.1361229Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:17.1361465Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1361634Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1361909Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1362571Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1363199Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1363828Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1364469Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1365099Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1365758Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1366384Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1367037Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1367667Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1368306Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1368437Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:17.1368510Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1368554Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1368592Z unimplemented []
2025-12-04T09:45:17.1368654Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1368755Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1369333Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1369372Z graph_break []
2025-12-04T09:45:17.1369447Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1369487Z Autotune Choices Stats:
2025-12-04T09:45:17.1370243Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.1370382Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1370539Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1370702Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1371338Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1371944Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1372559Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1373167Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1373775Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1374395Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1375004Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1375651Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1376258Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1376889Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1377019Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:17.1377061Z Autotune Choices Stats:
2025-12-04T09:45:17.1377832Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.1378053Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1378218Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1378510Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1379145Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1379789Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1380448Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1381086Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1381716Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1382343Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1382984Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1383605Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1384259Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1384884Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1385025Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:17.1385100Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1385142Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1385181Z unimplemented []
2025-12-04T09:45:17.1385241Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1385344Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1385936Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1385976Z graph_break []
2025-12-04T09:45:17.1386049Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1386089Z Autotune Choices Stats:
2025-12-04T09:45:17.1386839Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:17.1386977Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1387092Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1387251Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1387886Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1388498Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1389111Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1389725Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1390332Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1390974Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1391605Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1392208Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1392837Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1393439Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1393582Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:17.1393625Z Autotune Choices Stats:
2025-12-04T09:45:17.1394392Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:17.1394612Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1394779Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1395058Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1395708Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1396338Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1396985Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1397607Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1398245Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1398873Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1399504Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1400140Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1400802Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1401456Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1401586Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:17.1401660Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1401702Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1401740Z unimplemented []
2025-12-04T09:45:17.1401814Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1401916Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1402499Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1402539Z graph_break []
2025-12-04T09:45:17.1402614Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1402656Z Autotune Choices Stats:
2025-12-04T09:45:17.1403404Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.1403534Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1403647Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1403806Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1404434Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1405065Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1405673Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1406283Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1406901Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1407506Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1408111Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1408725Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1409350Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1409956Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1410086Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:17.1410126Z Autotune Choices Stats:
2025-12-04T09:45:17.1410928Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.1411162Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1411328Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1411613Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1412240Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1412886Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1413513Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1414167Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1414796Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1415429Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1416054Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1416680Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1417314Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1417957Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1418088Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:17.1418163Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1418205Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1418245Z unimplemented []
2025-12-04T09:45:17.1418304Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1418405Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1418981Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1419032Z graph_break []
2025-12-04T09:45:17.1419105Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1419146Z Autotune Choices Stats:
2025-12-04T09:45:17.1419898Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:17.1420027Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1420143Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1420302Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1420965Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1421587Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1422222Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1422833Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1423443Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1424058Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1424679Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1425285Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1425901Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1426521Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1426651Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:17.1426692Z Autotune Choices Stats:
2025-12-04T09:45:17.1427452Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:17.1427682Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1427848Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1428123Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1428757Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1429381Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1430015Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1430681Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1431315Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1431946Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1432580Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1433201Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1433825Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1434602Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1434747Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:17.1434821Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1434865Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1434904Z unimplemented []
2025-12-04T09:45:17.1434965Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1435076Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1435651Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1435689Z graph_break []
2025-12-04T09:45:17.1435766Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1435806Z Autotune Choices Stats:
2025-12-04T09:45:17.1436545Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.1436687Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1436801Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1436961Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1437573Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1438194Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1438802Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1439443Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1440046Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1440698Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1441331Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1441941Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1442541Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1443177Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1443328Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:17.1443368Z Autotune Choices Stats:
2025-12-04T09:45:17.1444155Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.1444374Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1444540Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1444847Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1445479Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1446105Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1446723Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1447363Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1448010Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1448638Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1449262Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1449900Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1450574Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1451223Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1451354Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:17.1451430Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1451472Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1451512Z unimplemented []
2025-12-04T09:45:17.1451571Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1451692Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1452293Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1452336Z graph_break []
2025-12-04T09:45:17.1452409Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1452450Z Autotune Choices Stats:
2025-12-04T09:45:17.1453197Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.1453355Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1453473Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1453632Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1454257Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1454870Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1455498Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1456104Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1456733Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1457338Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1457962Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1458574Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1459186Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1459803Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1459931Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:17.1459974Z Autotune Choices Stats:
2025-12-04T09:45:17.1460783Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.1461016Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1461183Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1461459Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1462088Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1462725Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1463351Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1463996Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1464625Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1465278Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1465906Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1466535Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1467172Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1467800Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1467928Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:17.1468004Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1468048Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1468086Z unimplemented []
2025-12-04T09:45:17.1468149Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1468260Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1468836Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1468884Z graph_break []
2025-12-04T09:45:17.1468957Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1468998Z Autotune Choices Stats:
2025-12-04T09:45:17.1469758Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:17.1469888Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1470002Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1470182Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1470812Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1471419Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1472032Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1472679Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1473283Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1473923Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1474532Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1475166Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1475777Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1476388Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1476519Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:17.1476558Z Autotune Choices Stats:
2025-12-04T09:45:17.1477337Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.1477566Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1477741Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1478023Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1478656Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1479296Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1479933Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1480590Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1481240Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1481867Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1482527Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1483155Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1483795Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1484416Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1484546Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:17.1484619Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1484663Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1484701Z unimplemented []
2025-12-04T09:45:17.1484764Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1484863Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1485453Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1485493Z graph_break []
2025-12-04T09:45:17.1485566Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1485607Z Autotune Choices Stats:
2025-12-04T09:45:17.1486357Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:17.1486506Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1486620Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1486781Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1487398Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1488017Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1488625Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1489235Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1489859Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1490514Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1491166Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1491782Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1492402Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1493006Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1493135Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:17.1493176Z Autotune Choices Stats:
2025-12-04T09:45:17.1493941Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.1494174Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1494342Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1494630Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1495270Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1495897Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1496534Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1497157Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1497782Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1498421Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1499042Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1499695Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1500322Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1501002Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1501130Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:17.1501205Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1501247Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1501286Z unimplemented []
2025-12-04T09:45:17.1501347Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1501449Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1502025Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1502064Z graph_break []
2025-12-04T09:45:17.1502138Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1502178Z Autotune Choices Stats:
2025-12-04T09:45:17.1502946Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.1503101Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1503214Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1503395Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1504015Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1504620Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1505250Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1505873Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1506477Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1507101Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1507712Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1508342Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1508946Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1509555Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1509688Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:17.1509728Z Autotune Choices Stats:
2025-12-04T09:45:17.1510531Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:17.1510752Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1510918Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1511217Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1511853Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1512528Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1513155Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1513792Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1514438Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1515082Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1515719Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1516345Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1516996Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1517623Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1517763Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:17.1517836Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1517881Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1517920Z unimplemented []
2025-12-04T09:45:17.1517981Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1518079Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1518654Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1518695Z graph_break []
2025-12-04T09:45:17.1518767Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1518808Z Autotune Choices Stats:
2025-12-04T09:45:17.1519568Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.1519699Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1519814Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1519975Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1520650Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1521259Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1521869Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1522493Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1523103Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1523724Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1524362Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1524987Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1525594Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1526212Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1526355Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:17.1526396Z Autotune Choices Stats:
2025-12-04T09:45:17.1527153Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.1527374Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1527541Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1527821Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1528466Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1529095Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1529743Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1530374Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1531051Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1531682Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1532310Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1532966Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1533617Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1534242Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1534373Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:17.1534448Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1534511Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1534551Z unimplemented []
2025-12-04T09:45:17.1534612Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1534713Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1535287Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1535325Z graph_break []
2025-12-04T09:45:17.1535399Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1535440Z Autotune Choices Stats:
2025-12-04T09:45:17.1536186Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.1536314Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1536429Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1536603Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1537217Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1537845Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1538454Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1539061Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1539679Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1540288Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1540932Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1541575Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1542227Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1542830Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1542960Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:17.1543001Z Autotune Choices Stats:
2025-12-04T09:45:17.1543786Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.1544002Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1544169Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1544450Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1545081Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1545731Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1546376Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1547003Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1547631Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1548270Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1548898Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1549524Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1550157Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1550856Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1550987Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:17.1551060Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1551104Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1551144Z unimplemented []
2025-12-04T09:45:17.1551206Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1551305Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1551872Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1551928Z graph_break []
2025-12-04T09:45:17.1552002Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1552043Z Autotune Choices Stats:
2025-12-04T09:45:17.1552795Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1708", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.1552925Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1553038Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1553201Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1553855Z   triton_flex_attention_1708 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1554460Z   triton_flex_attention_1706 0.0107 ms 93.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1555097Z   triton_flex_attention_1709 0.0110 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1555699Z   triton_flex_attention_1704 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1556305Z   triton_flex_attention_1707 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1556918Z   triton_flex_attention_1705 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1557540Z   triton_flex_attention_1724 0.0146 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1558170Z   triton_flex_attention_1716 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1558775Z   triton_flex_attention_1722 0.0160 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1559399Z   triton_flex_attention_1702 0.0166 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1559531Z SingleProcess AUTOTUNE benchmarking takes 0.1988 seconds and 0.5275 seconds precompiling for 24 choices
2025-12-04T09:45:17.1559573Z Autotune Choices Stats:
2025-12-04T09:45:17.1560339Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1743", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.1560590Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1560756Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1561032Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1561674Z   triton_flex_attention_backward_1743 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1562307Z   triton_flex_attention_backward_1737 0.0208 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1562951Z   triton_flex_attention_backward_1734 0.0213 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1563603Z   triton_flex_attention_backward_1735 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1564230Z   triton_flex_attention_backward_1745 0.0232 ms 76.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1564852Z   triton_flex_attention_backward_1744 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1565489Z   triton_flex_attention_backward_1742 0.0249 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1566122Z   triton_flex_attention_backward_1747 0.0252 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1566764Z   triton_flex_attention_backward_1738 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1567388Z   triton_flex_attention_backward_1729 0.0264 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1567535Z SingleProcess AUTOTUNE benchmarking takes 0.2428 seconds and 0.7372 seconds precompiling for 22 choices
2025-12-04T09:45:17.1567610Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1567653Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1567706Z unimplemented []
2025-12-04T09:45:17.1567768Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1567869Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1568441Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1568480Z graph_break []
2025-12-04T09:45:17.1568554Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1568606Z Autotune Choices Stats:
2025-12-04T09:45:17.1569341Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1754", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.1569468Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1569583Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1569748Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1570366Z   triton_flex_attention_1754 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1571038Z   triton_flex_attention_1752 0.0110 ms 90.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1571648Z   triton_flex_attention_1755 0.0114 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1572284Z   triton_flex_attention_1753 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1572891Z   triton_flex_attention_1750 0.0125 ms 79.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1573510Z   triton_flex_attention_1751 0.0143 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1574135Z   triton_flex_attention_1770 0.0149 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1574742Z   triton_flex_attention_1762 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1575362Z   triton_flex_attention_1768 0.0163 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1575963Z   triton_flex_attention_1748 0.0170 ms 58.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1576104Z SingleProcess AUTOTUNE benchmarking takes 0.2060 seconds and 0.4503 seconds precompiling for 24 choices
2025-12-04T09:45:17.1576146Z Autotune Choices Stats:
2025-12-04T09:45:17.1576932Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1789", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.1577150Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1577318Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1577607Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1578231Z   triton_flex_attention_backward_1789 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1578855Z   triton_flex_attention_backward_1783 0.0209 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1579507Z   triton_flex_attention_backward_1780 0.0216 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1580130Z   triton_flex_attention_backward_1781 0.0217 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1580839Z   triton_flex_attention_backward_1791 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1581470Z   triton_flex_attention_backward_1790 0.0235 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1582106Z   triton_flex_attention_backward_1788 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1582750Z   triton_flex_attention_backward_1793 0.0255 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1583374Z   triton_flex_attention_backward_1775 0.0264 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1584022Z   triton_flex_attention_backward_1784 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1584155Z SingleProcess AUTOTUNE benchmarking takes 0.2498 seconds and 0.6949 seconds precompiling for 22 choices
2025-12-04T09:45:17.1584228Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1584285Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1584323Z unimplemented []
2025-12-04T09:45:17.1584384Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1584483Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1585072Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1585113Z graph_break []
2025-12-04T09:45:17.1585186Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1585228Z Autotune Choices Stats:
2025-12-04T09:45:17.1585975Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1800", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.1586114Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1586227Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1586390Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1587010Z   triton_flex_attention_1800 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1587618Z   triton_flex_attention_1798 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1588244Z   triton_flex_attention_1801 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1588850Z   triton_flex_attention_1796 0.0121 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1589477Z   triton_flex_attention_1799 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1590085Z   triton_flex_attention_1816 0.0145 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1590736Z   triton_flex_attention_1797 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1591341Z   triton_flex_attention_1808 0.0152 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1591949Z   triton_flex_attention_1814 0.0161 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1592571Z   triton_flex_attention_1806 0.0168 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1592703Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.5450 seconds precompiling for 24 choices
2025-12-04T09:45:17.1592745Z Autotune Choices Stats:
2025-12-04T09:45:17.1593519Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1835", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.1593752Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1593918Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1594197Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1594848Z   triton_flex_attention_backward_1835 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1595476Z   triton_flex_attention_backward_1829 0.0210 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1596100Z   triton_flex_attention_backward_1826 0.0212 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1596731Z   triton_flex_attention_backward_1827 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1597368Z   triton_flex_attention_backward_1837 0.0231 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1598023Z   triton_flex_attention_backward_1836 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1598652Z   triton_flex_attention_backward_1839 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1599296Z   triton_flex_attention_backward_1834 0.0252 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1599927Z   triton_flex_attention_backward_1830 0.0260 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1600594Z   triton_flex_attention_backward_1821 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1600723Z SingleProcess AUTOTUNE benchmarking takes 0.2508 seconds and 0.7770 seconds precompiling for 22 choices
2025-12-04T09:45:17.1600798Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1600843Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1600882Z unimplemented []
2025-12-04T09:45:17.1600973Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1601074Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1601657Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1601710Z graph_break []
2025-12-04T09:45:17.1601785Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1601827Z Autotune Choices Stats:
2025-12-04T09:45:17.1602582Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1846", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.1602711Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1602828Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1603008Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1603629Z   triton_flex_attention_1846 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1604244Z   triton_flex_attention_1844 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1604857Z   triton_flex_attention_1847 0.0114 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1605473Z   triton_flex_attention_1842 0.0122 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1606084Z   triton_flex_attention_1845 0.0124 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1606709Z   triton_flex_attention_1843 0.0144 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1607321Z   triton_flex_attention_1862 0.0146 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1607940Z   triton_flex_attention_1854 0.0154 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1608548Z   triton_flex_attention_1860 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1609150Z   triton_flex_attention_1840 0.0167 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1609279Z SingleProcess AUTOTUNE benchmarking takes 0.2278 seconds and 0.3492 seconds precompiling for 24 choices
2025-12-04T09:45:17.1609322Z Autotune Choices Stats:
2025-12-04T09:45:17.1610099Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1881", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.1610325Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1610553Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1610833Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1611466Z   triton_flex_attention_backward_1881 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1612106Z   triton_flex_attention_backward_1875 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1612736Z   triton_flex_attention_backward_1873 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1613361Z   triton_flex_attention_backward_1872 0.0216 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1614010Z   triton_flex_attention_backward_1882 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1614636Z   triton_flex_attention_backward_1883 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1615292Z   triton_flex_attention_backward_1880 0.0254 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1615919Z   triton_flex_attention_backward_1885 0.0254 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1616555Z   triton_flex_attention_backward_1876 0.0263 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1617177Z   triton_flex_attention_backward_1867 0.0267 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1617306Z SingleProcess AUTOTUNE benchmarking takes 0.2409 seconds and 0.8665 seconds precompiling for 22 choices
2025-12-04T09:45:17.1617379Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1617425Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1617463Z unimplemented []
2025-12-04T09:45:17.1617524Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1617623Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1618210Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 74), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 28), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 12), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1618250Z graph_break []
2025-12-04T09:45:17.1618325Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1618377Z Autotune Choices Stats:
2025-12-04T09:45:17.1619138Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1892", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.1619268Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1619382Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1619544Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1620167Z   triton_flex_attention_1892 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1620811Z   triton_flex_attention_1890 0.0109 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1621422Z   triton_flex_attention_1893 0.0114 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1622046Z   triton_flex_attention_1888 0.0122 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1622680Z   triton_flex_attention_1891 0.0123 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1623319Z   triton_flex_attention_1889 0.0144 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1623933Z   triton_flex_attention_1908 0.0146 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1624559Z   triton_flex_attention_1900 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1625188Z   triton_flex_attention_1906 0.0161 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1625792Z   triton_flex_attention_1886 0.0167 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1625923Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3466 seconds precompiling for 24 choices
2025-12-04T09:45:17.1625963Z Autotune Choices Stats:
2025-12-04T09:45:17.1626748Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1927", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.1626968Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1627134Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1627423Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1628065Z   triton_flex_attention_backward_1927 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1628695Z   triton_flex_attention_backward_1921 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1629328Z   triton_flex_attention_backward_1918 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1629973Z   triton_flex_attention_backward_1919 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1630627Z   triton_flex_attention_backward_1929 0.0231 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1631278Z   triton_flex_attention_backward_1928 0.0233 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1631918Z   triton_flex_attention_backward_1926 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1632558Z   triton_flex_attention_backward_1931 0.0254 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1633189Z   triton_flex_attention_backward_1922 0.0261 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1633828Z   triton_flex_attention_backward_1913 0.0263 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1633956Z SingleProcess AUTOTUNE benchmarking takes 0.2431 seconds and 0.7860 seconds precompiling for 22 choices
2025-12-04T09:45:17.1634033Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1634076Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1634116Z unimplemented []
2025-12-04T09:45:17.1634176Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1634278Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1634847Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1634888Z graph_break []
2025-12-04T09:45:17.1634964Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1635007Z Autotune Choices Stats:
2025-12-04T09:45:17.1635769Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1938", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.1635906Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1636031Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1636194Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1636809Z   triton_flex_attention_1938 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1637437Z   triton_flex_attention_1936 0.0109 ms 93.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1638057Z   triton_flex_attention_1939 0.0116 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1638666Z   triton_flex_attention_1934 0.0122 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1639266Z   triton_flex_attention_1937 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1639884Z   triton_flex_attention_1935 0.0144 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1640563Z   triton_flex_attention_1954 0.0148 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1641169Z   triton_flex_attention_1946 0.0154 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1641775Z   triton_flex_attention_1952 0.0164 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1642396Z   triton_flex_attention_1944 0.0170 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1642526Z SingleProcess AUTOTUNE benchmarking takes 0.2077 seconds and 0.3245 seconds precompiling for 24 choices
2025-12-04T09:45:17.1642569Z Autotune Choices Stats:
2025-12-04T09:45:17.1643329Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1973", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.1643546Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1643726Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1644005Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1644646Z   triton_flex_attention_backward_1973 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1645287Z   triton_flex_attention_backward_1967 0.0211 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1645917Z   triton_flex_attention_backward_1965 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1646553Z   triton_flex_attention_backward_1964 0.0217 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1647187Z   triton_flex_attention_backward_1975 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1647819Z   triton_flex_attention_backward_1974 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1648466Z   triton_flex_attention_backward_1972 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1649119Z   triton_flex_attention_backward_1977 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1649744Z   triton_flex_attention_backward_1968 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1650368Z   triton_flex_attention_backward_1959 0.0266 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1650546Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.8096 seconds precompiling for 22 choices
2025-12-04T09:45:17.1650620Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1650667Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1650705Z unimplemented []
2025-12-04T09:45:17.1650766Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1650866Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1651448Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1651487Z graph_break []
2025-12-04T09:45:17.1651562Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1651604Z Autotune Choices Stats:
2025-12-04T09:45:17.1652358Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1984", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.1652487Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1652601Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1652776Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1653405Z   triton_flex_attention_1984 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1654011Z   triton_flex_attention_1982 0.0109 ms 95.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1654619Z   triton_flex_attention_1985 0.0113 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1655239Z   triton_flex_attention_1980 0.0122 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1655860Z   triton_flex_attention_1983 0.0124 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1656466Z   triton_flex_attention_1981 0.0142 ms 73.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1657085Z   triton_flex_attention_2000 0.0146 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1657715Z   triton_flex_attention_1992 0.0151 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1658320Z   triton_flex_attention_1998 0.0160 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1658925Z   triton_flex_attention_1978 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1659065Z SingleProcess AUTOTUNE benchmarking takes 0.2059 seconds and 0.3341 seconds precompiling for 24 choices
2025-12-04T09:45:17.1659106Z Autotune Choices Stats:
2025-12-04T09:45:17.1659866Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2019", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.1660083Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1660249Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1660555Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1661219Z   triton_flex_attention_backward_2019 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1661882Z   triton_flex_attention_backward_2013 0.0210 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1662512Z   triton_flex_attention_backward_2010 0.0214 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1663155Z   triton_flex_attention_backward_2011 0.0214 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1663808Z   triton_flex_attention_backward_2021 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1664444Z   triton_flex_attention_backward_2020 0.0233 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1665068Z   triton_flex_attention_backward_2018 0.0250 ms 72.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1665725Z   triton_flex_attention_backward_2023 0.0253 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1666371Z   triton_flex_attention_backward_2014 0.0262 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1666998Z   triton_flex_attention_backward_2005 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1667128Z SingleProcess AUTOTUNE benchmarking takes 0.2422 seconds and 0.7502 seconds precompiling for 22 choices
2025-12-04T09:45:17.1667234Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:17.1667283Z Traceback (most recent call last):
2025-12-04T09:45:17.1667441Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:17.1667484Z     self.assertTrue(
2025-12-04T09:45:17.1667595Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:17.1667646Z     raise self.failureException(msg)
2025-12-04T09:45:17.1667776Z AssertionError: False is not true : Log file /tmp/tmpz0t24o3o/flex_attention_configs.json was not created
2025-12-04T09:45:17.1667780Z 
2025-12-04T09:45:17.1667856Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.1668023Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.1668027Z 
2025-12-04T09:45:17.1668116Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.1668195Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1668238Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1668279Z unimplemented []
2025-12-04T09:45:17.1668340Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1668919Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:17.1669022Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1669060Z graph_break []
2025-12-04T09:45:17.1669136Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1669637Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:17.1669698Z   current_size = base.storage().size()
2025-12-04T09:45:17.1669740Z Autotune Choices Stats:
2025-12-04T09:45:17.1670538Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.1670669Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1670784Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1670949Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1671562Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1672183Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1672791Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1673396Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1674015Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1674636Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1675243Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1675851Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1676461Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1677062Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1677193Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:17.1677235Z Autotune Choices Stats:
2025-12-04T09:45:17.1678008Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.1678228Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1678393Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1678682Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1679324Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1679953Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1680604Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1681229Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1681868Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1682513Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1683158Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1683870Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1684498Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1685136Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1685268Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:17.1685346Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1685390Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1685429Z unimplemented []
2025-12-04T09:45:17.1685489Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1685590Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1686161Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1686202Z graph_break []
2025-12-04T09:45:17.1686278Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1686320Z Autotune Choices Stats:
2025-12-04T09:45:17.1687070Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.1687210Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1687336Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1687499Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1688111Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1688722Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1689335Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1689940Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1690578Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1691198Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1691823Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1692427Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1693029Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1693644Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1693774Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:17.1693817Z Autotune Choices Stats:
2025-12-04T09:45:17.1694571Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.1694792Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1694975Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1695251Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1695910Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1696536Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1697161Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1697794Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1698415Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1699039Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1699671Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1700318Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1700973Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1701594Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1701746Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:17.1701825Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1701869Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1701908Z unimplemented []
2025-12-04T09:45:17.1701969Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1702070Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1702652Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1702690Z graph_break []
2025-12-04T09:45:17.1702765Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1702806Z Autotune Choices Stats:
2025-12-04T09:45:17.1703561Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.1703690Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1703803Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1703980Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1704608Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1705208Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1705813Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1706427Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1707030Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1707636Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1708263Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1708893Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1709498Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1710096Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1710239Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:17.1710280Z Autotune Choices Stats:
2025-12-04T09:45:17.1711073Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.1711294Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1711459Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1711736Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1712403Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1713056Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1713684Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1714314Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1714956Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1715587Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1716216Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1716850Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1717500Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1718122Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1718252Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:17.1718338Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1718383Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1718423Z unimplemented []
2025-12-04T09:45:17.1718487Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1718585Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1719156Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1719197Z graph_break []
2025-12-04T09:45:17.1719271Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1719315Z Autotune Choices Stats:
2025-12-04T09:45:17.1720078Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:17.1720208Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1720324Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1720531Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1721148Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1721782Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1722392Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1723009Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1723615Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1724218Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1724839Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1725438Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1726064Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1726665Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1726796Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:17.1726854Z Autotune Choices Stats:
2025-12-04T09:45:17.1727614Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.1727834Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1728006Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1728282Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1728916Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1729573Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1730214Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1730856Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1731485Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1732130Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1732749Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1733378Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1734016Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1734663Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1734792Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:17.1734868Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1734912Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1734955Z unimplemented []
2025-12-04T09:45:17.1735015Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1735116Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1735698Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1735746Z graph_break []
2025-12-04T09:45:17.1735821Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1735863Z Autotune Choices Stats:
2025-12-04T09:45:17.1736607Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:17.1736735Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1736850Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1737016Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1737632Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1738238Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1738866Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1739470Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1740087Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1740727Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1741343Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1741976Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1742584Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1743224Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1743360Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:17.1743402Z Autotune Choices Stats:
2025-12-04T09:45:17.1744179Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.1744410Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1744580Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1744862Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1745497Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1746123Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1746759Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1747406Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1748034Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1748665Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1749299Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1749931Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1750635Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1751259Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1751401Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:17.1751476Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1751533Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1751573Z unimplemented []
2025-12-04T09:45:17.1751636Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1751735Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1752315Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1752355Z graph_break []
2025-12-04T09:45:17.1752429Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1752488Z Autotune Choices Stats:
2025-12-04T09:45:17.1753231Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:17.1753360Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1753477Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1753639Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1754249Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1754859Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1755464Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1756088Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1756692Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1757309Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1757917Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1758523Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1759138Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1759739Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1759890Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:17.1759934Z Autotune Choices Stats:
2025-12-04T09:45:17.1760735Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.1760955Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1761136Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1761414Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1762040Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1762668Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1763304Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1763931Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1764581Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1765212Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1765843Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1766473Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1767093Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1767728Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1767856Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:17.1767931Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1767983Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1768023Z unimplemented []
2025-12-04T09:45:17.1768083Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1768186Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1768776Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1768817Z graph_break []
2025-12-04T09:45:17.1768892Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1768933Z Autotune Choices Stats:
2025-12-04T09:45:17.1769670Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:17.1769807Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1769924Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1770089Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1770735Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1771337Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1771958Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1772561Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1773191Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1773794Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1774410Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1775013Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1775617Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1776240Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1776372Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:17.1776412Z Autotune Choices Stats:
2025-12-04T09:45:17.1777194Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.1777414Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1777581Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1777859Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1778501Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1779144Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1779768Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1780402Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1781064Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1781717Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1782343Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1782986Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1783611Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1784254Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1784391Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:17.1784467Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1784511Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1784549Z unimplemented []
2025-12-04T09:45:17.1784627Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1784727Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1785306Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1785356Z graph_break []
2025-12-04T09:45:17.1785429Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1785473Z Autotune Choices Stats:
2025-12-04T09:45:17.1786229Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:17.1786361Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1786475Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1786648Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1787259Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1787858Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1788466Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1789084Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1789686Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1790310Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1790949Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1791578Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1792181Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1792785Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1792916Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:17.1792959Z Autotune Choices Stats:
2025-12-04T09:45:17.1793825Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.1794057Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1794238Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1794516Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1795149Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1795786Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1796414Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1797037Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1797676Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1798306Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1798948Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1799578Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1800216Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1800873Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1801002Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:17.1801076Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1801120Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1801158Z unimplemented []
2025-12-04T09:45:17.1801218Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1801319Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1801908Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1801948Z graph_break []
2025-12-04T09:45:17.1802022Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1802064Z Autotune Choices Stats:
2025-12-04T09:45:17.1802838Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.1802966Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1803080Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1803242Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1803851Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1804462Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1805073Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1805674Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1806291Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1806925Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1807540Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1808145Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1808760Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1809365Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1809496Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:17.1809539Z Autotune Choices Stats:
2025-12-04T09:45:17.1810318Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.1810563Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1810730Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1811025Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1811673Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1812300Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1812937Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1813556Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1814188Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1814834Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1815458Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1816112Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1816739Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1817366Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1817496Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:17.1817571Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1817617Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1817655Z unimplemented []
2025-12-04T09:45:17.1817720Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1817819Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1818397Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1818435Z graph_break []
2025-12-04T09:45:17.1818510Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1818551Z Autotune Choices Stats:
2025-12-04T09:45:17.1819306Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:17.1819445Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1819559Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1819732Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1820356Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1820996Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1821614Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1822220Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1822827Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1823446Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1824085Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1824690Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1825298Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1825907Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1826038Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:17.1826079Z Autotune Choices Stats:
2025-12-04T09:45:17.1826833Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.1827052Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1827229Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1827507Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1828142Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1828787Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1829415Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1830048Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1830700Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1831324Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1831969Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1832620Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1833247Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1833871Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1834010Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:17.1835946Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1835995Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1836036Z unimplemented []
2025-12-04T09:45:17.1836100Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1836205Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1836787Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1836827Z graph_break []
2025-12-04T09:45:17.1836902Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1836946Z Autotune Choices Stats:
2025-12-04T09:45:17.1837709Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:17.1837839Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1837958Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1838132Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1838763Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1839374Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1839984Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1840627Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1841226Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1841827Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1842484Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1843115Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1843722Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1844328Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1844475Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:17.1844519Z Autotune Choices Stats:
2025-12-04T09:45:17.1845302Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.1845525Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1845694Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1845976Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1846624Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1847267Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1847895Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1848518Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1849159Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1849784Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1850430Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1851077Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1851733Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1852364Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1852498Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:17.1852574Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1852635Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1852675Z unimplemented []
2025-12-04T09:45:17.1852739Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1852839Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1853422Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1853460Z graph_break []
2025-12-04T09:45:17.1853537Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1853580Z Autotune Choices Stats:
2025-12-04T09:45:17.1854324Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.1854453Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1854567Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1854746Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1855362Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1855988Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1856596Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1857202Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1857826Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1858441Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1859053Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1859666Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1860290Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1860930Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1861061Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:17.1861115Z Autotune Choices Stats:
2025-12-04T09:45:17.1861869Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.1862088Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1862255Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1862538Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1863187Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1863832Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1864485Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1865112Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1865755Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1866386Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1867014Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1867650Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1868298Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1868944Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1869075Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:17.1869153Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1869196Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1869237Z unimplemented []
2025-12-04T09:45:17.1869298Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1869399Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1869974Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1870023Z graph_break []
2025-12-04T09:45:17.1870097Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1870139Z Autotune Choices Stats:
2025-12-04T09:45:17.1870918Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.1871047Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1871162Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1871325Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1871976Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1872583Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1873221Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1873823Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1874443Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1875047Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1875655Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1876274Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1876879Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1877505Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1877635Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:17.1877677Z Autotune Choices Stats:
2025-12-04T09:45:17.1878440Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.1878670Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1878838Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1879117Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1879758Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1880384Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1881063Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1881709Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1882341Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1882974Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1883608Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1884238Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1884874Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1885503Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1885641Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:17.1885717Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1885762Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1885809Z unimplemented []
2025-12-04T09:45:17.1885873Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1885972Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1886549Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1886587Z graph_break []
2025-12-04T09:45:17.1886662Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1886714Z Autotune Choices Stats:
2025-12-04T09:45:17.1887451Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.1887579Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1887692Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1887857Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1888474Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1889094Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1889700Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1890333Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1890953Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1891571Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1892178Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1892786Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1893421Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1894028Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1894171Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:17.1894213Z Autotune Choices Stats:
2025-12-04T09:45:17.1894990Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.1895208Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1895374Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1895662Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1896306Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1896953Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1897605Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1898231Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1898879Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1899507Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1900128Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1900800Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1901429Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1902091Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1902222Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:17.1902297Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1902341Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1902394Z unimplemented []
2025-12-04T09:45:17.1902456Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1902555Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1903150Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1903190Z graph_break []
2025-12-04T09:45:17.1903263Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1903306Z Autotune Choices Stats:
2025-12-04T09:45:17.1904050Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:17.1904193Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1904308Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1904469Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1905092Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1905700Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1906319Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1906918Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1907539Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1908145Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1908755Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1909356Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1909962Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1910607Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1910736Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:17.1910778Z Autotune Choices Stats:
2025-12-04T09:45:17.1911550Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:17.1911782Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1911949Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1912228Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1912868Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1913505Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1914138Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1914790Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1915414Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1916066Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1916691Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1917332Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1917967Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1918590Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1918720Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:17.1918797Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1918839Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1918879Z unimplemented []
2025-12-04T09:45:17.1918953Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1919054Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1919629Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1919680Z graph_break []
2025-12-04T09:45:17.1919755Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1919796Z Autotune Choices Stats:
2025-12-04T09:45:17.1920590Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.1920720Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1920835Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1921017Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1921630Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1922238Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1922844Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1923471Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1924075Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1924708Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1925319Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1925930Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1926531Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1927132Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1927264Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:17.1927304Z Autotune Choices Stats:
2025-12-04T09:45:17.1928082Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.1928309Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1928486Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1928764Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1929396Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1930033Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1930700Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1931325Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1931972Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1932601Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1933245Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1933877Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1934523Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1935148Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1935280Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:17.1935355Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1935399Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1935438Z unimplemented []
2025-12-04T09:45:17.1935499Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1935598Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1936182Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.1936222Z graph_break []
2025-12-04T09:45:17.1936295Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1936338Z Autotune Choices Stats:
2025-12-04T09:45:17.1937095Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.1937241Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1937356Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1937516Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1938126Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1938735Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1939344Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1939958Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1940636Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1941238Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1941875Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1942477Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1943097Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1943695Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1943827Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:17.1943872Z Autotune Choices Stats:
2025-12-04T09:45:17.1944640Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.1944861Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1945029Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1945320Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1945962Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1946589Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1947230Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1947863Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1948498Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1949155Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1949780Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1950470Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1951098Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1951735Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1951865Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:17.1951943Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1951986Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1952026Z unimplemented []
2025-12-04T09:45:17.1952088Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1952189Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1952764Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1952803Z graph_break []
2025-12-04T09:45:17.1952878Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1952919Z Autotune Choices Stats:
2025-12-04T09:45:17.1953678Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:17.1953824Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1953938Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1954113Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1954732Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1955336Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1955954Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1956563Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1957161Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1957770Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1958398Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1959010Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1959622Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1960234Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1960365Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:17.1960446Z Autotune Choices Stats:
2025-12-04T09:45:17.1961208Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:17.1961424Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1961593Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1961891Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1962520Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1963172Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1963797Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1964431Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1965057Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1965688Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1966316Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1966944Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1967592Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1968220Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1968361Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:17.1968434Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1968477Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1968516Z unimplemented []
2025-12-04T09:45:17.1968577Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1968675Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1969244Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1969283Z graph_break []
2025-12-04T09:45:17.1969357Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1969399Z Autotune Choices Stats:
2025-12-04T09:45:17.1970160Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.1970291Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1970421Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1970585Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1971227Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1971828Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1972432Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1973047Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1973651Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1974252Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1974868Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1975494Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1976101Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1976706Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1976849Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:17.1976891Z Autotune Choices Stats:
2025-12-04T09:45:17.1977653Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:17.1977874Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1978040Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1978316Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1978968Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1979595Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1980237Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1980879Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1981514Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1982145Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1982768Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1983414Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1984068Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1984695Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1984826Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:17.1984901Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.1984967Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.1985007Z unimplemented []
2025-12-04T09:45:17.1985068Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.1985169Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.1985750Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.1985789Z graph_break []
2025-12-04T09:45:17.1985864Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.1985908Z Autotune Choices Stats:
2025-12-04T09:45:17.1986655Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.1986783Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1986897Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1987061Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1987685Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1988315Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1988921Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1989527Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1990135Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1990779Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.1991388Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1992007Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1992638Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1993246Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1993375Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:17.1993418Z Autotune Choices Stats:
2025-12-04T09:45:17.1994192Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.1994412Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.1994579Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.1994860Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.1995494Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1996140Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1996786Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1997407Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1998054Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1998694Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1999324Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.1999956Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2000627Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2001275Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2001405Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:17.2001480Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2001523Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2001562Z unimplemented []
2025-12-04T09:45:17.2001623Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2001723Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2002296Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2002345Z graph_break []
2025-12-04T09:45:17.2002420Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2002461Z Autotune Choices Stats:
2025-12-04T09:45:17.2003205Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.2003334Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2003448Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2003611Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2004239Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2004841Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2005468Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2006075Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2006679Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2007289Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2007895Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2008504Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2009119Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2009740Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2009872Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:17.2009912Z Autotune Choices Stats:
2025-12-04T09:45:17.2010694Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.2010928Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2011093Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2011375Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2012008Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2012637Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2013287Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2013953Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2014584Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2015217Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2015850Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2016480Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2017108Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2017749Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2017892Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:17.2017968Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2018011Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2018050Z unimplemented []
2025-12-04T09:45:17.2018121Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2018223Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2018802Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2018842Z graph_break []
2025-12-04T09:45:17.2018915Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2018957Z Autotune Choices Stats:
2025-12-04T09:45:17.2019704Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:17.2019831Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2019947Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2020108Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2020748Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2021372Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2021974Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2022601Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2023204Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2023825Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2024431Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2025047Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2025671Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2026275Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2026414Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:17.2026457Z Autotune Choices Stats:
2025-12-04T09:45:17.2027234Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.2027451Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2027622Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2027914Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2028543Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2029174Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2029798Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2030489Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2031145Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2031776Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2032408Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2033058Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2033689Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2034346Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2034477Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:17.2034552Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2034595Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2034643Z unimplemented []
2025-12-04T09:45:17.2034705Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2034803Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2035391Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2035429Z graph_break []
2025-12-04T09:45:17.2035503Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2035543Z Autotune Choices Stats:
2025-12-04T09:45:17.2036285Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.2036426Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2036539Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2036702Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2037326Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2037935Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2038559Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2039159Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2039785Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2040392Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2041049Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2041657Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2042272Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2042898Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2043030Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:17.2043071Z Autotune Choices Stats:
2025-12-04T09:45:17.2043842Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.2044074Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2044238Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2044520Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2045153Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2045794Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2046427Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2047069Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2047699Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2048351Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2048978Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2049625Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2050256Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2050916Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2051047Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:17.2051123Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2051166Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2051206Z unimplemented []
2025-12-04T09:45:17.2051288Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2051390Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2051963Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2052015Z graph_break []
2025-12-04T09:45:17.2052088Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2052132Z Autotune Choices Stats:
2025-12-04T09:45:17.2052892Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.2053020Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2053135Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2053309Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2053925Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2054536Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2055146Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2055769Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2056373Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2057001Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2057611Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2058232Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2058861Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2059472Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2059603Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:17.2059647Z Autotune Choices Stats:
2025-12-04T09:45:17.2060477Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.2060710Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2060892Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2061169Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2061806Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2062448Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2063093Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2063723Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2064370Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2064999Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2065649Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2066281Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2066921Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2067563Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2067693Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:17.2067768Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2067811Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2067850Z unimplemented []
2025-12-04T09:45:17.2067911Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2068010Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2068594Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2068633Z graph_break []
2025-12-04T09:45:17.2068707Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2068747Z Autotune Choices Stats:
2025-12-04T09:45:17.2069518Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.2069647Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2069760Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2069926Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2070589Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2071222Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2071848Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2072458Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2073078Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2073684Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2074313Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2074922Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2075537Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2076166Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2076298Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:17.2076338Z Autotune Choices Stats:
2025-12-04T09:45:17.2077117Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.2077338Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2077503Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2077793Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2078441Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2079066Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2079706Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2080333Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2081020Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2081684Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2082310Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2083002Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2083631Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2084275Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2084406Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:17.2084482Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2084525Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2084565Z unimplemented []
2025-12-04T09:45:17.2084626Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2084726Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2085309Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2085348Z graph_break []
2025-12-04T09:45:17.2085421Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2085466Z Autotune Choices Stats:
2025-12-04T09:45:17.2086226Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:17.2086362Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2086478Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2086651Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2087267Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2087886Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2088510Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2089121Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2089730Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2090354Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2091036Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2091643Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2092256Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2092874Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2093005Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:17.2093081Z Autotune Choices Stats:
2025-12-04T09:45:17.2093948Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:17.2094231Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2094427Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2094721Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2095457Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2096124Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2096747Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2097401Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2098034Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2098662Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2099312Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2099959Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2100623Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2101246Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2101391Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:17.2101467Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2101512Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2101550Z unimplemented []
2025-12-04T09:45:17.2101611Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2101711Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2102287Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2102326Z graph_break []
2025-12-04T09:45:17.2102400Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2102442Z Autotune Choices Stats:
2025-12-04T09:45:17.2103220Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.2103350Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2103465Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2103640Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2104272Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2104875Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2105499Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2106110Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2106721Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2107325Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2107943Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2108572Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2109179Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2109785Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2109925Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:17.2109968Z Autotune Choices Stats:
2025-12-04T09:45:17.2110764Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.2110983Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2111151Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2111428Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2112089Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2112738Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2113357Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2113976Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2114633Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2115277Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2115912Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2116552Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2117191Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2117819Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2117946Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:17.2118022Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2118077Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2118115Z unimplemented []
2025-12-04T09:45:17.2118177Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2118276Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2118850Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2118889Z graph_break []
2025-12-04T09:45:17.2118963Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2119003Z Autotune Choices Stats:
2025-12-04T09:45:17.2119747Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:17.2119874Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2119991Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2120164Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2120818Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2121453Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2122060Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2122683Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2123300Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2123913Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2124528Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2125143Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2125769Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2126366Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2126496Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:17.2126546Z Autotune Choices Stats:
2025-12-04T09:45:17.2127309Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:17.2127527Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2127697Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2127976Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2128608Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2129258Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2129908Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2130561Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2131189Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2131824Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2132453Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2133083Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2133729Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2134378Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2134509Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:17.2134583Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2134627Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2134665Z unimplemented []
2025-12-04T09:45:17.2134726Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2134827Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2135405Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2135456Z graph_break []
2025-12-04T09:45:17.2135530Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2135574Z Autotune Choices Stats:
2025-12-04T09:45:17.2136335Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.2136463Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2136577Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2136738Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2137367Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2137971Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2138596Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2139200Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2139813Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2140454Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2141063Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2141690Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2142294Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2142922Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2143053Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:17.2143094Z Autotune Choices Stats:
2025-12-04T09:45:17.2143866Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.2144099Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2144264Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2144547Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2145174Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2145810Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2146435Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2148132Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2148760Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2150124Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2150793Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2151419Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2152042Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2152666Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2152814Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:17.2152894Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2152937Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2152976Z unimplemented []
2025-12-04T09:45:17.2153037Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2153175Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2153747Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2153787Z graph_break []
2025-12-04T09:45:17.2153875Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2153917Z Autotune Choices Stats:
2025-12-04T09:45:17.2154666Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:17.2154796Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2154913Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2155076Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2155700Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2156310Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2156921Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2157558Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2158160Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2158786Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2159393Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2159994Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2160631Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2161239Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2161381Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:17.2161426Z Autotune Choices Stats:
2025-12-04T09:45:17.2162207Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:17.2162428Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2162615Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2162910Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2163544Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2164170Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2164797Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2165440Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2166090Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2166717Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2167359Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2167982Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2168606Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2169234Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2169363Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:17.2169436Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2169490Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2169528Z unimplemented []
2025-12-04T09:45:17.2169591Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2169690Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2170269Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2170308Z graph_break []
2025-12-04T09:45:17.2170384Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2170467Z Autotune Choices Stats:
2025-12-04T09:45:17.2171221Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.2171376Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2171504Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2171666Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2172278Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2172910Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2173542Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2174162Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2174791Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2175392Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2176024Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2176628Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2177230Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2177834Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2177968Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:17.2178011Z Autotune Choices Stats:
2025-12-04T09:45:17.2178783Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.2179027Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2179194Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2179473Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2180134Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2180795Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2181413Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2182035Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2182662Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2183320Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2183938Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2184594Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2185218Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2185843Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2185973Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:17.2186049Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2186092Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2186130Z unimplemented []
2025-12-04T09:45:17.2186191Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2186295Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2186872Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2186923Z graph_break []
2025-12-04T09:45:17.2186997Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2187039Z Autotune Choices Stats:
2025-12-04T09:45:17.2187793Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.2187922Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2188048Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2188209Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2188836Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2189441Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2190046Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2190691Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2191312Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2191929Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2192535Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2193173Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2193781Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2194386Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2194515Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:17.2194558Z Autotune Choices Stats:
2025-12-04T09:45:17.2195318Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.2195555Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2195722Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2196018Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2196651Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2197299Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2197919Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2198542Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2199162Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2199786Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2200492Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2201114Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2201782Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2202406Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2202537Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:17.2202612Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2202654Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2202693Z unimplemented []
2025-12-04T09:45:17.2202752Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2202853Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2203433Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2203470Z graph_break []
2025-12-04T09:45:17.2203545Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2203612Z Autotune Choices Stats:
2025-12-04T09:45:17.2204360Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:17.2204502Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2204618Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2204783Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2205413Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2206038Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2206646Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2207266Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2207862Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2208478Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2209101Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2209703Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2210334Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2210968Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2211099Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:17.2211141Z Autotune Choices Stats:
2025-12-04T09:45:17.2211903Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.2212123Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2212288Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2212591Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2213245Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2213869Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2214526Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2215157Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2215792Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2216426Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2217057Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2217693Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2218320Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2218962Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2219093Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:17.2219168Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2219214Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2219251Z unimplemented []
2025-12-04T09:45:17.2219311Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2219412Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2219980Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2220018Z graph_break []
2025-12-04T09:45:17.2220092Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2220131Z Autotune Choices Stats:
2025-12-04T09:45:17.2220900Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:17.2221048Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2221163Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2221322Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2221956Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2222563Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2223202Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2223808Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2224411Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2225020Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2225631Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2226249Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2226857Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2227489Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2227620Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:17.2227661Z Autotune Choices Stats:
2025-12-04T09:45:17.2228413Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.2228631Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2228798Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2229081Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2229722Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2230355Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2231004Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2231658Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2232286Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2232924Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2233552Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2234188Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2234823Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2235447Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2235599Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:17.2235675Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2235717Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2235757Z unimplemented []
2025-12-04T09:45:17.2235817Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2235918Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2236489Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2236529Z graph_break []
2025-12-04T09:45:17.2236602Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2236645Z Autotune Choices Stats:
2025-12-04T09:45:17.2237389Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.2237515Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2237631Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2237802Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2238425Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2239034Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2239662Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2240268Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2240903Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2241505Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2242116Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2242738Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2243368Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2243969Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2244127Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:17.2244171Z Autotune Choices Stats:
2025-12-04T09:45:17.2244934Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:17.2245152Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2245319Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2245595Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2246231Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2246867Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2247516Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2248138Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2248787Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2249419Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2250054Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2250728Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2251383Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2252009Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2252139Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:17.2252226Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2252271Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2252310Z unimplemented []
2025-12-04T09:45:17.2252372Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2252472Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2253055Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2253094Z graph_break []
2025-12-04T09:45:17.2253168Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2253209Z Autotune Choices Stats:
2025-12-04T09:45:17.2253953Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.2254079Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2254194Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2254355Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2254966Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2255597Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2256207Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2256831Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2257433Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2258036Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2258658Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2259279Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2259910Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2260553Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2260707Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:17.2260749Z Autotune Choices Stats:
2025-12-04T09:45:17.2261548Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.2261768Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2261935Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2262219Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2262862Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2263492Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2264138Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2264763Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2265423Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2266058Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2266682Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2267313Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2267956Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2268591Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2268722Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:17.2268799Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2268841Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2268880Z unimplemented []
2025-12-04T09:45:17.2268940Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2269041Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2269624Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2269676Z graph_break []
2025-12-04T09:45:17.2269750Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2269792Z Autotune Choices Stats:
2025-12-04T09:45:17.2270560Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.2270691Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2270807Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2270966Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2271581Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2272185Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2272820Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2273425Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2274060Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2274662Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2275274Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2275887Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2276499Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2277124Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2277254Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:17.2277297Z Autotune Choices Stats:
2025-12-04T09:45:17.2278056Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.2278295Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2278464Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2278741Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2279369Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2279998Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2280748Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2281405Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2282035Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2282690Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2283320Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2283946Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2284576Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2285206Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2285347Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:17.2285420Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2285463Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2285500Z unimplemented []
2025-12-04T09:45:17.2285573Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2285672Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2286248Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2286297Z graph_break []
2025-12-04T09:45:17.2286372Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2286413Z Autotune Choices Stats:
2025-12-04T09:45:17.2287169Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1708", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.2287296Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2287411Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2287577Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2288195Z   triton_flex_attention_1708 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2288801Z   triton_flex_attention_1706 0.0107 ms 93.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2289426Z   triton_flex_attention_1709 0.0110 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2290056Z   triton_flex_attention_1704 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2290716Z   triton_flex_attention_1707 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2291358Z   triton_flex_attention_1705 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2291967Z   triton_flex_attention_1724 0.0146 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2292590Z   triton_flex_attention_1716 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2293188Z   triton_flex_attention_1722 0.0160 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2293796Z   triton_flex_attention_1702 0.0166 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2293939Z SingleProcess AUTOTUNE benchmarking takes 0.1988 seconds and 0.5275 seconds precompiling for 24 choices
2025-12-04T09:45:17.2293982Z Autotune Choices Stats:
2025-12-04T09:45:17.2294776Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1743", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.2294996Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2295173Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2295460Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2296094Z   triton_flex_attention_backward_1743 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2296740Z   triton_flex_attention_backward_1737 0.0208 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2297373Z   triton_flex_attention_backward_1734 0.0213 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2298006Z   triton_flex_attention_backward_1735 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2298659Z   triton_flex_attention_backward_1745 0.0232 ms 76.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2299281Z   triton_flex_attention_backward_1744 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2299927Z   triton_flex_attention_backward_1742 0.0249 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2300590Z   triton_flex_attention_backward_1747 0.0252 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2301221Z   triton_flex_attention_backward_1738 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2301841Z   triton_flex_attention_backward_1729 0.0264 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2301972Z SingleProcess AUTOTUNE benchmarking takes 0.2428 seconds and 0.7372 seconds precompiling for 22 choices
2025-12-04T09:45:17.2302065Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2302107Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2302147Z unimplemented []
2025-12-04T09:45:17.2302206Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2302307Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2302895Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2302937Z graph_break []
2025-12-04T09:45:17.2303010Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2303051Z Autotune Choices Stats:
2025-12-04T09:45:17.2303788Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1754", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.2303940Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2304058Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2304218Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2304833Z   triton_flex_attention_1754 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2305442Z   triton_flex_attention_1752 0.0110 ms 90.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2306048Z   triton_flex_attention_1755 0.0114 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2306665Z   triton_flex_attention_1753 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2307286Z   triton_flex_attention_1750 0.0125 ms 79.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2307887Z   triton_flex_attention_1751 0.0143 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2308513Z   triton_flex_attention_1770 0.0149 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2309116Z   triton_flex_attention_1762 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2309721Z   triton_flex_attention_1768 0.0163 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2310323Z   triton_flex_attention_1748 0.0170 ms 58.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2310477Z SingleProcess AUTOTUNE benchmarking takes 0.2060 seconds and 0.4503 seconds precompiling for 24 choices
2025-12-04T09:45:17.2310543Z Autotune Choices Stats:
2025-12-04T09:45:17.2311308Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1789", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.2311529Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2311697Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2311976Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2312641Z   triton_flex_attention_backward_1789 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2313267Z   triton_flex_attention_backward_1783 0.0209 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2313888Z   triton_flex_attention_backward_1780 0.0216 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2314514Z   triton_flex_attention_backward_1781 0.0217 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2315135Z   triton_flex_attention_backward_1791 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2315784Z   triton_flex_attention_backward_1790 0.0235 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2316404Z   triton_flex_attention_backward_1788 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2317049Z   triton_flex_attention_backward_1793 0.0255 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2317678Z   triton_flex_attention_backward_1775 0.0264 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2318302Z   triton_flex_attention_backward_1784 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2318431Z SingleProcess AUTOTUNE benchmarking takes 0.2498 seconds and 0.6949 seconds precompiling for 22 choices
2025-12-04T09:45:17.2318506Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2318548Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2318587Z unimplemented []
2025-12-04T09:45:17.2318648Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2318749Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2319334Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2319372Z graph_break []
2025-12-04T09:45:17.2319449Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2319490Z Autotune Choices Stats:
2025-12-04T09:45:17.2320238Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1800", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.2320367Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2320528Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2320691Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2321323Z   triton_flex_attention_1800 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2321928Z   triton_flex_attention_1798 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2322556Z   triton_flex_attention_1801 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2323186Z   triton_flex_attention_1796 0.0121 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2323806Z   triton_flex_attention_1799 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2324427Z   triton_flex_attention_1816 0.0145 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2325032Z   triton_flex_attention_1797 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2325658Z   triton_flex_attention_1808 0.0152 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2326262Z   triton_flex_attention_1814 0.0161 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2326863Z   triton_flex_attention_1806 0.0168 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2326993Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.5450 seconds precompiling for 24 choices
2025-12-04T09:45:17.2327034Z Autotune Choices Stats:
2025-12-04T09:45:17.2327797Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1835", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.2328027Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2328204Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2328485Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2329118Z   triton_flex_attention_backward_1835 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2329765Z   triton_flex_attention_backward_1829 0.0210 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2330393Z   triton_flex_attention_backward_1826 0.0212 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2331048Z   triton_flex_attention_backward_1827 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2331690Z   triton_flex_attention_backward_1837 0.0231 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2332330Z   triton_flex_attention_backward_1836 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2332969Z   triton_flex_attention_backward_1839 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2333594Z   triton_flex_attention_backward_1834 0.0252 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2334245Z   triton_flex_attention_backward_1830 0.0260 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2334869Z   triton_flex_attention_backward_1821 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2334998Z SingleProcess AUTOTUNE benchmarking takes 0.2508 seconds and 0.7770 seconds precompiling for 22 choices
2025-12-04T09:45:17.2335073Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2335116Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2335153Z unimplemented []
2025-12-04T09:45:17.2335215Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2335314Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2335884Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2335933Z graph_break []
2025-12-04T09:45:17.2336006Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2336048Z Autotune Choices Stats:
2025-12-04T09:45:17.2336795Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1846", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.2336923Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2337038Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2337197Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2337808Z   triton_flex_attention_1846 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2338441Z   triton_flex_attention_1844 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2339045Z   triton_flex_attention_1847 0.0114 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2339651Z   triton_flex_attention_1842 0.0122 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2340249Z   triton_flex_attention_1845 0.0124 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2340929Z   triton_flex_attention_1843 0.0144 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2344048Z   triton_flex_attention_1862 0.0146 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2344662Z   triton_flex_attention_1854 0.0154 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2345299Z   triton_flex_attention_1860 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2345908Z   triton_flex_attention_1840 0.0167 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2346041Z SingleProcess AUTOTUNE benchmarking takes 0.2278 seconds and 0.3492 seconds precompiling for 24 choices
2025-12-04T09:45:17.2346085Z Autotune Choices Stats:
2025-12-04T09:45:17.2346844Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1881", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.2347066Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2347253Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2347532Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2348186Z   triton_flex_attention_backward_1881 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2348813Z   triton_flex_attention_backward_1875 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2349460Z   triton_flex_attention_backward_1873 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2350088Z   triton_flex_attention_backward_1872 0.0216 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2350761Z   triton_flex_attention_backward_1882 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2351392Z   triton_flex_attention_backward_1883 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2352034Z   triton_flex_attention_backward_1880 0.0254 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2352694Z   triton_flex_attention_backward_1885 0.0254 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2353322Z   triton_flex_attention_backward_1876 0.0263 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2353971Z   triton_flex_attention_backward_1867 0.0267 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2354101Z SingleProcess AUTOTUNE benchmarking takes 0.2409 seconds and 0.8665 seconds precompiling for 22 choices
2025-12-04T09:45:17.2354184Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2354228Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2354270Z unimplemented []
2025-12-04T09:45:17.2354334Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2354439Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2355029Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 74), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 28), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 12), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2355068Z graph_break []
2025-12-04T09:45:17.2355146Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2355187Z Autotune Choices Stats:
2025-12-04T09:45:17.2355955Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1892", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.2356095Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2356211Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2356387Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2357007Z   triton_flex_attention_1892 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2357640Z   triton_flex_attention_1890 0.0109 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2358245Z   triton_flex_attention_1893 0.0114 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2358864Z   triton_flex_attention_1888 0.0122 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2359467Z   triton_flex_attention_1891 0.0123 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2360071Z   triton_flex_attention_1889 0.0144 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2360748Z   triton_flex_attention_1908 0.0146 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2361351Z   triton_flex_attention_1900 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2361957Z   triton_flex_attention_1906 0.0161 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2362594Z   triton_flex_attention_1886 0.0167 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2362725Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3466 seconds precompiling for 24 choices
2025-12-04T09:45:17.2362766Z Autotune Choices Stats:
2025-12-04T09:45:17.2363521Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1927", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.2363740Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2363909Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2364190Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2364847Z   triton_flex_attention_backward_1927 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2365485Z   triton_flex_attention_backward_1921 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2366115Z   triton_flex_attention_backward_1918 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2366759Z   triton_flex_attention_backward_1919 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2367391Z   triton_flex_attention_backward_1929 0.0231 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2368024Z   triton_flex_attention_backward_1928 0.0233 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2368651Z   triton_flex_attention_backward_1926 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2369299Z   triton_flex_attention_backward_1931 0.0254 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2369926Z   triton_flex_attention_backward_1922 0.0261 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2370618Z   triton_flex_attention_backward_1913 0.0263 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2370751Z SingleProcess AUTOTUNE benchmarking takes 0.2431 seconds and 0.7860 seconds precompiling for 22 choices
2025-12-04T09:45:17.2370827Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2370872Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2370909Z unimplemented []
2025-12-04T09:45:17.2370972Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2371073Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2371647Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2371685Z graph_break []
2025-12-04T09:45:17.2371759Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2371802Z Autotune Choices Stats:
2025-12-04T09:45:17.2372538Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1938", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.2372668Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2372796Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2372958Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2373598Z   triton_flex_attention_1938 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2374200Z   triton_flex_attention_1936 0.0109 ms 93.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2374837Z   triton_flex_attention_1939 0.0116 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2375443Z   triton_flex_attention_1934 0.0122 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2376051Z   triton_flex_attention_1937 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2376670Z   triton_flex_attention_1935 0.0144 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2377282Z   triton_flex_attention_1954 0.0148 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2377904Z   triton_flex_attention_1946 0.0154 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2378514Z   triton_flex_attention_1952 0.0164 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2379142Z   triton_flex_attention_1944 0.0170 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2379273Z SingleProcess AUTOTUNE benchmarking takes 0.2077 seconds and 0.3245 seconds precompiling for 24 choices
2025-12-04T09:45:17.2379315Z Autotune Choices Stats:
2025-12-04T09:45:17.2380070Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1973", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.2380289Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2380487Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2380769Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2381403Z   triton_flex_attention_backward_1973 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2382064Z   triton_flex_attention_backward_1967 0.0211 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2382686Z   triton_flex_attention_backward_1965 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2383339Z   triton_flex_attention_backward_1964 0.0217 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2383969Z   triton_flex_attention_backward_1975 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2384601Z   triton_flex_attention_backward_1974 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2385223Z   triton_flex_attention_backward_1972 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2385854Z   triton_flex_attention_backward_1977 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2386501Z   triton_flex_attention_backward_1968 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2387124Z   triton_flex_attention_backward_1959 0.0266 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2387263Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.8096 seconds precompiling for 22 choices
2025-12-04T09:45:17.2387338Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2387382Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2387421Z unimplemented []
2025-12-04T09:45:17.2387482Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2387592Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2388165Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2388205Z graph_break []
2025-12-04T09:45:17.2388279Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2388321Z Autotune Choices Stats:
2025-12-04T09:45:17.2389064Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1984", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.2389192Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2389308Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2389470Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2390089Z   triton_flex_attention_1984 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2390766Z   triton_flex_attention_1982 0.0109 ms 95.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2391371Z   triton_flex_attention_1985 0.0113 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2392005Z   triton_flex_attention_1980 0.0122 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2392609Z   triton_flex_attention_1983 0.0124 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2393212Z   triton_flex_attention_1981 0.0142 ms 73.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2393820Z   triton_flex_attention_2000 0.0146 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2394424Z   triton_flex_attention_1992 0.0151 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2395054Z   triton_flex_attention_1998 0.0160 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2395653Z   triton_flex_attention_1978 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2395792Z SingleProcess AUTOTUNE benchmarking takes 0.2059 seconds and 0.3341 seconds precompiling for 24 choices
2025-12-04T09:45:17.2395834Z Autotune Choices Stats:
2025-12-04T09:45:17.2396601Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2019", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.2396824Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2396993Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2397275Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2397908Z   triton_flex_attention_backward_2019 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2398528Z   triton_flex_attention_backward_2013 0.0210 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2399175Z   triton_flex_attention_backward_2010 0.0214 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2399795Z   triton_flex_attention_backward_2011 0.0214 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2400476Z   triton_flex_attention_backward_2021 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2401106Z   triton_flex_attention_backward_2020 0.0233 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2401733Z   triton_flex_attention_backward_2018 0.0250 ms 72.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2402381Z   triton_flex_attention_backward_2023 0.0253 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2403007Z   triton_flex_attention_backward_2014 0.0262 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2403660Z   triton_flex_attention_backward_2005 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2403791Z SingleProcess AUTOTUNE benchmarking takes 0.2422 seconds and 0.7502 seconds precompiling for 22 choices
2025-12-04T09:45:17.2403866Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2403910Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2403947Z unimplemented []
2025-12-04T09:45:17.2404008Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2404120Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2404708Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2404747Z graph_break []
2025-12-04T09:45:17.2404821Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2404862Z Autotune Choices Stats:
2025-12-04T09:45:17.2405610Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2030", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.2405739Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2405853Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2406018Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2406640Z   triton_flex_attention_2030 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2407253Z   triton_flex_attention_2028 0.0109 ms 94.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2407874Z   triton_flex_attention_2031 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2408480Z   triton_flex_attention_2026 0.0126 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2409104Z   triton_flex_attention_2029 0.0127 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2409701Z   triton_flex_attention_2027 0.0142 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2410312Z   triton_flex_attention_2046 0.0147 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2410959Z   triton_flex_attention_2038 0.0152 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2411565Z   triton_flex_attention_2044 0.0162 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2412196Z   triton_flex_attention_2024 0.0165 ms 62.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2412328Z SingleProcess AUTOTUNE benchmarking takes 0.2047 seconds and 0.3631 seconds precompiling for 24 choices
2025-12-04T09:45:17.2412370Z Autotune Choices Stats:
2025-12-04T09:45:17.2413133Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2065", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017799999564886093, "best_triton_pos": 0}
2025-12-04T09:45:17.2413378Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2413545Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2413826Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2414463Z   triton_flex_attention_backward_2065 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2415093Z   triton_flex_attention_backward_2059 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2415711Z   triton_flex_attention_backward_2056 0.0213 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2416355Z   triton_flex_attention_backward_2057 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2416982Z   triton_flex_attention_backward_2067 0.0230 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2417628Z   triton_flex_attention_backward_2066 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2418250Z   triton_flex_attention_backward_2064 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2418890Z   triton_flex_attention_backward_2069 0.0252 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2419530Z   triton_flex_attention_backward_2060 0.0260 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2420156Z   triton_flex_attention_backward_2051 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2420296Z SingleProcess AUTOTUNE benchmarking takes 0.2494 seconds and 0.8153 seconds precompiling for 22 choices
2025-12-04T09:45:17.2420391Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:17.2420490Z Traceback (most recent call last):
2025-12-04T09:45:17.2420650Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:17.2420691Z     self.assertTrue(
2025-12-04T09:45:17.2420800Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:17.2420850Z     raise self.failureException(msg)
2025-12-04T09:45:17.2420978Z AssertionError: False is not true : Log file /tmp/tmpr50o_zw3/flex_attention_configs.json was not created
2025-12-04T09:45:17.2420982Z 
2025-12-04T09:45:17.2421060Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.2421245Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.2421247Z 
2025-12-04T09:45:17.2421342Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.2421418Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2421463Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2421501Z unimplemented []
2025-12-04T09:45:17.2421577Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2422150Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:17.2422251Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2422288Z graph_break []
2025-12-04T09:45:17.2422364Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2422856Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:17.2422907Z   current_size = base.storage().size()
2025-12-04T09:45:17.2422947Z Autotune Choices Stats:
2025-12-04T09:45:17.2423690Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.2423835Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2423950Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2424110Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2424740Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2425341Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2425966Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2426566Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2427181Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2427796Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2428399Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2429028Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2429631Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2430257Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2430390Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:17.2430462Z Autotune Choices Stats:
2025-12-04T09:45:17.2431224Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.2431444Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2431610Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2431887Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2432517Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2433170Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2433793Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2434441Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2435067Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2435692Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2436316Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2436951Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2437607Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2438225Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2438366Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:17.2438443Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2438486Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2438525Z unimplemented []
2025-12-04T09:45:17.2438596Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2438697Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2439272Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2439312Z graph_break []
2025-12-04T09:45:17.2439385Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2439428Z Autotune Choices Stats:
2025-12-04T09:45:17.2440172Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.2440302Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2440462Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2440624Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2441238Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2441869Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2442471Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2443090Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2443816Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2444416Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2445016Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2445621Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2446250Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2446850Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2446991Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:17.2447033Z Autotune Choices Stats:
2025-12-04T09:45:17.2447806Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.2448024Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2448196Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2448475Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2449116Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2449741Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2450382Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2451035Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2451689Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2452315Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2452948Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2453589Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2454220Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2454872Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2455004Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:17.2455079Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2455123Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2455161Z unimplemented []
2025-12-04T09:45:17.2455224Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2455334Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2455920Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2455958Z graph_break []
2025-12-04T09:45:17.2456033Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2456073Z Autotune Choices Stats:
2025-12-04T09:45:17.2456808Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.2456936Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2457051Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2457214Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2457833Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2458448Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2459081Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2459684Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2460303Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2460935Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2461532Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2462136Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2462757Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2463389Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2463519Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:17.2463559Z Autotune Choices Stats:
2025-12-04T09:45:17.2464323Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.2464565Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2464732Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2465011Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2465648Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2466275Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2466892Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2467537Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2468158Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2468812Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2469439Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2470061Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2470723Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2471345Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2471489Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:17.2471564Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2471618Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2471657Z unimplemented []
2025-12-04T09:45:17.2471718Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2471821Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2472394Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2472448Z graph_break []
2025-12-04T09:45:17.2472525Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2472570Z Autotune Choices Stats:
2025-12-04T09:45:17.2473324Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:17.2473453Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2473573Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2473738Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2474346Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2474957Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2475580Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2476197Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2476793Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2477421Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2478027Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2478637Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2479243Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2479846Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2479988Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:17.2480030Z Autotune Choices Stats:
2025-12-04T09:45:17.2480821Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.2481054Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2481220Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2481519Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2482155Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2482782Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2483404Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2484025Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2484684Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2485308Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2485956Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2486584Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2487230Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2487868Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2488010Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:17.2488084Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2488128Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2488167Z unimplemented []
2025-12-04T09:45:17.2488228Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2488329Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2488920Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2488959Z graph_break []
2025-12-04T09:45:17.2489034Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2489073Z Autotune Choices Stats:
2025-12-04T09:45:17.2489814Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:17.2489963Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2490079Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2490241Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2490895Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2491503Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2492109Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2492720Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2493335Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2493934Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2494564Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2495170Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2495792Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2496397Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2496537Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:17.2496578Z Autotune Choices Stats:
2025-12-04T09:45:17.2497353Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.2497574Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2497740Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2498026Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2498669Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2499291Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2499918Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2500597Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2501239Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2501890Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2502511Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2503161Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2503783Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2504406Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2504537Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:17.2504612Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2504654Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2504693Z unimplemented []
2025-12-04T09:45:17.2504754Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2504857Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2505442Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2505482Z graph_break []
2025-12-04T09:45:17.2505556Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2505598Z Autotune Choices Stats:
2025-12-04T09:45:17.2506350Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:17.2506489Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2506604Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2506769Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2507398Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2508003Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2508607Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2509211Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2509824Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2510476Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2511085Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2511833Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2512434Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2513033Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2513162Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:17.2513204Z Autotune Choices Stats:
2025-12-04T09:45:17.2513957Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.2514197Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2514373Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2514653Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2515291Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2515929Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2516548Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2517165Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2517814Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2518453Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2519103Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2519734Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2520371Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2521045Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2521174Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:17.2521248Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2521291Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2521329Z unimplemented []
2025-12-04T09:45:17.2521390Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2521490Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2522056Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2522107Z graph_break []
2025-12-04T09:45:17.2522182Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2522222Z Autotune Choices Stats:
2025-12-04T09:45:17.2522959Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:17.2523089Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2523202Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2523364Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2523999Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2524603Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2525205Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2525811Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2526433Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2527046Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2527669Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2528268Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2528887Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2529486Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2529619Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:17.2529659Z Autotune Choices Stats:
2025-12-04T09:45:17.2530456Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.2530673Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2530860Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2531139Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2531770Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2532391Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2533032Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2533651Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2534274Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2534914Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2535542Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2536172Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2536793Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2537430Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2537560Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:17.2537635Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2537678Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2537719Z unimplemented []
2025-12-04T09:45:17.2537780Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2537881Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2538447Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2538486Z graph_break []
2025-12-04T09:45:17.2538558Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2538599Z Autotune Choices Stats:
2025-12-04T09:45:17.2539350Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:17.2539489Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2539605Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2539777Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2540388Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2541040Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2541642Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2542245Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2542845Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2543449Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2544071Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2544675Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2545275Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2545886Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2546015Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:17.2546056Z Autotune Choices Stats:
2025-12-04T09:45:17.2546815Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.2547033Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2547201Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2547486Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2548121Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2548752Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2549372Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2550019Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2550675Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2551295Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2551918Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2552567Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2553186Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2553817Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2553958Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:17.2554034Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2554077Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2554114Z unimplemented []
2025-12-04T09:45:17.2554175Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2554276Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2554840Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2554878Z graph_break []
2025-12-04T09:45:17.2554954Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2554994Z Autotune Choices Stats:
2025-12-04T09:45:17.2555728Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.2555856Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2555981Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2556142Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2556765Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2557367Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2557979Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2558578Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2559176Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2559768Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2560376Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2561027Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2561629Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2562248Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2562376Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:17.2562417Z Autotune Choices Stats:
2025-12-04T09:45:17.2563163Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.2563381Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2563546Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2563820Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2564450Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2565101Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2565718Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2566349Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2566979Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2567601Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2568233Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2568863Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2569499Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2570118Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2570257Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:17.2570332Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2570375Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2570450Z unimplemented []
2025-12-04T09:45:17.2570512Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2570626Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2571198Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2571236Z graph_break []
2025-12-04T09:45:17.2571310Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2571351Z Autotune Choices Stats:
2025-12-04T09:45:17.2572098Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:17.2572226Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2572341Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2572499Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2573116Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2573756Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2574428Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2575055Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2575654Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2576256Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2576857Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2577467Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2578092Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2578693Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2578832Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:17.2578872Z Autotune Choices Stats:
2025-12-04T09:45:17.2579635Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.2579853Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2580020Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2580298Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2580961Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2581584Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2582236Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2582855Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2583505Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2584128Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2584746Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2585389Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2586035Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2586677Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2586805Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:17.2586881Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2586923Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2586962Z unimplemented []
2025-12-04T09:45:17.2587021Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2587121Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2587714Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2587753Z graph_break []
2025-12-04T09:45:17.2587828Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2587868Z Autotune Choices Stats:
2025-12-04T09:45:17.2588615Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:17.2588744Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2588857Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2589019Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2589626Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2590222Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2590889Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2591485Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2592119Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2592715Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2593319Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2593919Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2594538Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2595178Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2595308Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:17.2595348Z Autotune Choices Stats:
2025-12-04T09:45:17.2596104Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.2596351Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2596517Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2596795Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2597425Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2598046Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2598684Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2599323Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2599948Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2600632Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2601251Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2601875Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2602494Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2603117Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2603259Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:17.2603333Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2603377Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2603432Z unimplemented []
2025-12-04T09:45:17.2603495Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2603594Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2604163Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2604222Z graph_break []
2025-12-04T09:45:17.2604295Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2604337Z Autotune Choices Stats:
2025-12-04T09:45:17.2605076Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.2605207Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2605322Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2605484Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2606106Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2606708Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2607316Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2607943Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2608543Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2609159Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2609762Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2610381Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2611005Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2611602Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2611747Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:17.2611788Z Autotune Choices Stats:
2025-12-04T09:45:17.2612565Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.2612782Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2612970Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2613258Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2613886Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2614509Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2615133Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2615753Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2616402Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2617033Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2617674Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2618300Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2618927Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2619548Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2619680Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:17.2619766Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2619808Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2619848Z unimplemented []
2025-12-04T09:45:17.2619909Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2620010Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2620645Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2620685Z graph_break []
2025-12-04T09:45:17.2620760Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2620800Z Autotune Choices Stats:
2025-12-04T09:45:17.2621537Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.2621697Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2621814Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2621976Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2622588Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2623180Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2623786Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2624400Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2625031Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2625630Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2626255Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2626857Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2627459Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2628066Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2628196Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:17.2628248Z Autotune Choices Stats:
2025-12-04T09:45:17.2629025Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.2629242Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2629409Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2629690Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2630336Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2630988Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2631608Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2632231Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2632852Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2633512Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2634130Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2634787Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2635410Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2636030Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2636160Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:17.2636235Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2636278Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2636316Z unimplemented []
2025-12-04T09:45:17.2636379Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2636479Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2637056Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2637095Z graph_break []
2025-12-04T09:45:17.2637169Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2637210Z Autotune Choices Stats:
2025-12-04T09:45:17.2637957Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.2638086Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2638211Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2638375Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2638997Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2639603Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2640208Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2640834Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2641437Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2642043Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2642645Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2643273Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2643871Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2644469Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2644601Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:17.2644643Z Autotune Choices Stats:
2025-12-04T09:45:17.2645396Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.2645628Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2645804Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2646082Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2646715Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2647355Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2647975Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2648596Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2649247Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2649884Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2650566Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2651186Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2651842Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2652460Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2652589Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:17.2652666Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2652708Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2652747Z unimplemented []
2025-12-04T09:45:17.2652807Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2652908Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2653475Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2653528Z graph_break []
2025-12-04T09:45:17.2653603Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2653643Z Autotune Choices Stats:
2025-12-04T09:45:17.2654391Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:17.2654519Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2654635Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2654794Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2655413Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2656036Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2656639Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2657233Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2657837Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2658445Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2659061Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2659667Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2660300Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2660938Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2661067Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:17.2661109Z Autotune Choices Stats:
2025-12-04T09:45:17.2661870Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:17.2662087Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2662267Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2662542Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2663189Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2663815Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2664458Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2665075Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2665701Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2666344Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2666972Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2667606Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2668228Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2668866Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2668996Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:17.2669069Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2669113Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2669150Z unimplemented []
2025-12-04T09:45:17.2669213Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2669314Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2669877Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2669915Z graph_break []
2025-12-04T09:45:17.2669990Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2670030Z Autotune Choices Stats:
2025-12-04T09:45:17.2670801Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.2670946Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2671061Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2671238Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2671846Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2672476Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2673078Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2673679Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2674296Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2674898Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2675509Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2676132Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2676733Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2677358Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2677487Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:17.2677527Z Autotune Choices Stats:
2025-12-04T09:45:17.2678282Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.2678499Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2678664Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2678939Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2679581Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2680222Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2680880Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2681528Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2682154Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2682776Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2683396Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2684032Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2684673Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2685283Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2685432Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:17.2685514Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2685556Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2685597Z unimplemented []
2025-12-04T09:45:17.2685657Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2685758Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2686335Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2686375Z graph_break []
2025-12-04T09:45:17.2686447Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2686488Z Autotune Choices Stats:
2025-12-04T09:45:17.2687232Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.2687361Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2687487Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2687648Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2688271Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2688872Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2689492Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2690090Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2690732Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2691327Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2691927Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2692562Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2693165Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2693763Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2693914Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:17.2693956Z Autotune Choices Stats:
2025-12-04T09:45:17.2694710Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.2694928Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2695096Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2695369Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2695995Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2696630Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2697290Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2697907Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2698551Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2699175Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2699794Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2700434Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2701091Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2701711Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2701852Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:17.2701927Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2701971Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2702010Z unimplemented []
2025-12-04T09:45:17.2702072Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2702189Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2702758Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2702796Z graph_break []
2025-12-04T09:45:17.2702874Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2702914Z Autotune Choices Stats:
2025-12-04T09:45:17.2703656Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:17.2703785Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2703899Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2704059Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2704668Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2705294Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2705890Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2706513Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2707110Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2707710Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2708314Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2708923Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2709549Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2710148Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2710290Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:17.2710332Z Autotune Choices Stats:
2025-12-04T09:45:17.2711152Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:17.2711370Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2711536Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2711814Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2712442Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2713073Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2713716Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2714336Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2714985Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2715606Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2716225Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2716846Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2717469Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2718108Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2718238Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:17.2718313Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2718355Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2718393Z unimplemented []
2025-12-04T09:45:17.2718453Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2718553Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2719146Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2719186Z graph_break []
2025-12-04T09:45:17.2719260Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2719302Z Autotune Choices Stats:
2025-12-04T09:45:17.2720056Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.2720185Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2720300Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2720482Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2721093Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2721695Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2722320Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2722918Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2723546Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2724139Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2724732Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2725338Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2725942Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2726559Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2726688Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:17.2726731Z Autotune Choices Stats:
2025-12-04T09:45:17.2727487Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:17.2727726Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2727893Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2728172Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2728806Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2729446Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2730068Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2730728Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2731351Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2732009Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2732626Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2733247Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2733884Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2734507Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2734651Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:17.2734725Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2734768Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2734808Z unimplemented []
2025-12-04T09:45:17.2734881Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2734980Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2735551Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2735600Z graph_break []
2025-12-04T09:45:17.2735676Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2735716Z Autotune Choices Stats:
2025-12-04T09:45:17.2736468Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.2736596Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2736711Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2736873Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2737486Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2738090Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2738691Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2739310Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2739898Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2740550Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2741148Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2741753Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2742358Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2742963Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2743106Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:17.2743146Z Autotune Choices Stats:
2025-12-04T09:45:17.2743923Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.2744140Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2744318Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2744607Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2745231Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2745853Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2746472Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2747091Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2747738Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2748360Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2748995Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2749620Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2750263Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2750912Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2751043Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:17.2751135Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2751177Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2751216Z unimplemented []
2025-12-04T09:45:17.2751277Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2751380Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2751970Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2752010Z graph_break []
2025-12-04T09:45:17.2752084Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2752129Z Autotune Choices Stats:
2025-12-04T09:45:17.2752876Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.2753031Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2753148Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2753308Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2753918Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2754513Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2755122Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2755723Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2756336Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2756934Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2757557Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2758162Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2758759Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2759362Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2759491Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:17.2759544Z Autotune Choices Stats:
2025-12-04T09:45:17.2760309Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.2760573Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2760741Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2761015Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2761690Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2762311Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2762928Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2763572Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2764192Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2764838Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2765458Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2766100Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2766716Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2767334Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2767461Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:17.2767537Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2767578Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2767618Z unimplemented []
2025-12-04T09:45:17.2767679Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2767779Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2768356Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2768405Z graph_break []
2025-12-04T09:45:17.2768479Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2768518Z Autotune Choices Stats:
2025-12-04T09:45:17.2769271Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:17.2769399Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2769523Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2769685Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2770307Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2770923Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2771522Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2772112Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2772732Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2773337Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2773938Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2774565Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2775158Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2775749Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2775878Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:17.2775918Z Autotune Choices Stats:
2025-12-04T09:45:17.2776682Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.2776911Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2777085Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2777364Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2778001Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2778651Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2779271Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2779891Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2780540Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2781181Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2781840Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2782462Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2783106Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2783729Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2783859Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:17.2783935Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2783979Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2784016Z unimplemented []
2025-12-04T09:45:17.2784078Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2784175Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2784749Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2784788Z graph_break []
2025-12-04T09:45:17.2784871Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2784913Z Autotune Choices Stats:
2025-12-04T09:45:17.2785663Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.2785794Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2785909Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2786070Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2786689Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2787322Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2787928Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2788523Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2789139Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2789749Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2790364Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2791009Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2791636Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2792241Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2792370Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:17.2792412Z Autotune Choices Stats:
2025-12-04T09:45:17.2793176Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.2793394Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2793574Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2793853Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2794504Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2795126Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2795764Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2796386Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2797037Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2797662Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2798288Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2798925Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2799552Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2800191Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2800321Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:17.2800397Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2800474Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2800514Z unimplemented []
2025-12-04T09:45:17.2800575Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2800677Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2801252Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2801290Z graph_break []
2025-12-04T09:45:17.2801367Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2801407Z Autotune Choices Stats:
2025-12-04T09:45:17.2802151Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.2802297Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2802413Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2802586Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2803198Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2803818Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2804433Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2805025Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2805623Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2806224Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2806837Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2807467Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2808072Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2808691Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2808822Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:17.2808863Z Autotune Choices Stats:
2025-12-04T09:45:17.2809617Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.2809831Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2810001Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2810283Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2810945Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2811579Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2812201Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2812846Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2813469Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2814093Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2814734Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2815367Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2815997Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2816617Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2816768Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:17.2816842Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2816886Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2816924Z unimplemented []
2025-12-04T09:45:17.2816986Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2817085Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2817657Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2817697Z graph_break []
2025-12-04T09:45:17.2817770Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2817812Z Autotune Choices Stats:
2025-12-04T09:45:17.2818552Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.2818682Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2818809Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2818971Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2819588Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2820195Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2820836Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2821440Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2822060Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2822661Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2823268Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2823899Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2824496Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2825111Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2825240Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:17.2825283Z Autotune Choices Stats:
2025-12-04T09:45:17.2826049Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.2826268Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2826435Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2826713Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2827365Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2827997Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2828647Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2829269Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2829912Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2830576Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2831200Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2831840Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2832580Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2833199Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2833338Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:17.2833414Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2833456Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2833495Z unimplemented []
2025-12-04T09:45:17.2833556Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2833668Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2834237Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2834279Z graph_break []
2025-12-04T09:45:17.2834353Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2834393Z Autotune Choices Stats:
2025-12-04T09:45:17.2835133Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:17.2835260Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2835375Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2835539Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2836151Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2836780Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2837385Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2838005Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2838605Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2839210Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2839813Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2840444Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2841064Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2841669Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2841811Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:17.2841852Z Autotune Choices Stats:
2025-12-04T09:45:17.2842622Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:17.2842839Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2843007Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2843284Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2843912Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2844527Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2845163Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2845790Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2846426Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2847051Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2849519Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2850148Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2850808Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2851464Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2851596Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:17.2851675Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2851721Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2851760Z unimplemented []
2025-12-04T09:45:17.2851824Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2851926Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2852545Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2852586Z graph_break []
2025-12-04T09:45:17.2852662Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2852704Z Autotune Choices Stats:
2025-12-04T09:45:17.2853449Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.2853580Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2853696Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2853858Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2854470Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2855073Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2855688Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2856289Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2856912Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2857504Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2858110Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2858713Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2859323Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2859944Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2860075Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:17.2860117Z Autotune Choices Stats:
2025-12-04T09:45:17.2860908Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.2861157Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2861326Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2861605Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2862241Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2862861Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2863494Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2864138Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2864754Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2865402Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2866024Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2866639Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2867264Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2867902Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2868039Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:17.2868116Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2868160Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2868209Z unimplemented []
2025-12-04T09:45:17.2868272Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2868374Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2868940Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2868996Z graph_break []
2025-12-04T09:45:17.2869071Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2869111Z Autotune Choices Stats:
2025-12-04T09:45:17.2869865Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:17.2869996Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2870112Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2870272Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2870926Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2871529Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2872155Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2872763Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2873362Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2874002Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2874603Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2875204Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2875811Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2876420Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2876564Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:17.2876607Z Autotune Choices Stats:
2025-12-04T09:45:17.2877373Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:17.2877599Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2877767Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2878053Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2878687Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2879312Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2879941Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2880598Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2881266Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2881882Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2882541Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2883163Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2883785Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2884399Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2884538Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:17.2884612Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2884655Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2884696Z unimplemented []
2025-12-04T09:45:17.2884758Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2884860Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2885441Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2885481Z graph_break []
2025-12-04T09:45:17.2885556Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2885597Z Autotune Choices Stats:
2025-12-04T09:45:17.2886335Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.2886483Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2886599Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2886759Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2887374Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2887978Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2888574Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2889185Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2889799Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2890398Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2891069Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2891669Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2892284Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2892884Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2893026Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:17.2893066Z Autotune Choices Stats:
2025-12-04T09:45:17.2893824Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.2894043Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2894210Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2894500Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2895141Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2895761Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2896384Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2897009Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2897641Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2898275Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2898895Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2899744Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2900372Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2901018Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2901146Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:17.2901221Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2901264Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2901303Z unimplemented []
2025-12-04T09:45:17.2901364Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2901466Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2902062Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2902101Z graph_break []
2025-12-04T09:45:17.2902174Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2902216Z Autotune Choices Stats:
2025-12-04T09:45:17.2902981Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:17.2903122Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2903238Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2903400Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2904023Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2904626Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2905228Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2905833Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2906442Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2907052Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2907661Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2908274Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2908875Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2909477Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2909610Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:17.2909659Z Autotune Choices Stats:
2025-12-04T09:45:17.2910458Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:17.2910694Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2910871Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2911156Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2911895Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2912585Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2913215Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2913845Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2914472Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2915113Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2915745Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2916375Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2917032Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2917668Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2917798Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:17.2917873Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2917916Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2917963Z unimplemented []
2025-12-04T09:45:17.2918025Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2918135Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2918708Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2918756Z graph_break []
2025-12-04T09:45:17.2918832Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2918873Z Autotune Choices Stats:
2025-12-04T09:45:17.2919629Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.2919759Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2919873Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2920034Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2920696Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2921300Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2921907Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2922509Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2923108Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2923730Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2924327Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2924928Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2925544Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2926144Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2926275Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:17.2926315Z Autotune Choices Stats:
2025-12-04T09:45:17.2927063Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.2927281Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2927456Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2927736Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2928376Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2928994Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2929638Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2930260Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2930939Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2931560Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2932197Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2932838Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2933457Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2934106Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2934234Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:17.2934311Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2934354Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2934393Z unimplemented []
2025-12-04T09:45:17.2934453Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2934557Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2935130Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2935170Z graph_break []
2025-12-04T09:45:17.2935243Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2935286Z Autotune Choices Stats:
2025-12-04T09:45:17.2936023Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.2936160Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2936286Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2936445Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2937050Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2937673Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2938266Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2938871Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2939483Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2940091Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2940751Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2941354Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2941977Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2942578Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2942706Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:17.2942749Z Autotune Choices Stats:
2025-12-04T09:45:17.2943511Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.2943728Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2943895Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2944174Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2944823Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2945446Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2946077Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2946701Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2947319Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2947947Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2948573Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2949214Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2949841Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2950503Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2950634Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:17.2950708Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2950751Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2950789Z unimplemented []
2025-12-04T09:45:17.2950853Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2950954Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2951532Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2951570Z graph_break []
2025-12-04T09:45:17.2951644Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2951684Z Autotune Choices Stats:
2025-12-04T09:45:17.2952422Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:17.2952562Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2952676Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2952837Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2953462Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2954064Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2954690Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2955282Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2955881Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2956485Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2957093Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2957705Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2958308Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2958929Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2959062Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:17.2959102Z Autotune Choices Stats:
2025-12-04T09:45:17.2959857Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.2960074Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2960239Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2960566Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2961204Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2961848Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2962468Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2963110Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2963734Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2964376Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2964995Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2965627Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2966278Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2966895Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2967039Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:17.2967114Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2967156Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2967195Z unimplemented []
2025-12-04T09:45:17.2967267Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2967369Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2967946Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.2967984Z graph_break []
2025-12-04T09:45:17.2968058Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2968100Z Autotune Choices Stats:
2025-12-04T09:45:17.2968840Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:17.2968970Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2969091Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2969251Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2969873Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2970523Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2971114Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2971741Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2972342Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2972946Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2973548Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2974155Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2974784Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2975377Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2975519Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:17.2975562Z Autotune Choices Stats:
2025-12-04T09:45:17.2976335Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.2976553Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2976719Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2976997Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2977624Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2978249Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2978889Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2979505Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2980155Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2980813Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2981430Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2982058Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2982690Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2983349Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2983478Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:17.2983551Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.2983595Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.2983635Z unimplemented []
2025-12-04T09:45:17.2983709Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.2983809Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.2984385Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.2984424Z graph_break []
2025-12-04T09:45:17.2984499Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.2984540Z Autotune Choices Stats:
2025-12-04T09:45:17.2985276Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.2985404Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2985517Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2985677Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2986287Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2986901Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2987514Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2988117Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2988735Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2989344Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2989949Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2990597Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2991209Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2991825Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2991957Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:17.2991997Z Autotune Choices Stats:
2025-12-04T09:45:17.2992778Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:17.2993006Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.2993172Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.2993452Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.2994076Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2994696Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2995320Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2995963Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2996590Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2997237Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2997858Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2998490Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.2999114Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2999747Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.2999877Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:17.2999964Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3000007Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3000045Z unimplemented []
2025-12-04T09:45:17.3000106Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3000207Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3000813Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3000871Z graph_break []
2025-12-04T09:45:17.3000945Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3000988Z Autotune Choices Stats:
2025-12-04T09:45:17.3001743Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.3001873Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3001988Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3002150Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3002772Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3003377Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3003989Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3004606Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3005200Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3005815Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3006418Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3007021Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3007622Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3008230Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3008362Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:17.3008413Z Autotune Choices Stats:
2025-12-04T09:45:17.3009169Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.3009397Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3009573Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3009849Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3010517Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3011141Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3011764Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3012397Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3013031Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3013663Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3014306Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3014928Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3015551Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3016172Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3016310Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:17.3016384Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3016428Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3016468Z unimplemented []
2025-12-04T09:45:17.3016531Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3016630Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3017212Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3017250Z graph_break []
2025-12-04T09:45:17.3017325Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3017365Z Autotune Choices Stats:
2025-12-04T09:45:17.3018116Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.3018255Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3018369Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3018531Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3019154Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3019764Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3020380Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3021021Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3021632Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3022236Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3022871Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3023473Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3024075Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3024679Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3024819Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:17.3024860Z Autotune Choices Stats:
2025-12-04T09:45:17.3025617Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.3025834Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3025998Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3026281Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3026920Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3027540Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3028162Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3028794Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3029432Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3030061Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3030712Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3031356Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3031981Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3032597Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3032727Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:17.3032802Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3032844Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3032884Z unimplemented []
2025-12-04T09:45:17.3032944Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3033057Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3033637Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3033676Z graph_break []
2025-12-04T09:45:17.3034150Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3034193Z Autotune Choices Stats:
2025-12-04T09:45:17.3034935Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1708", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.3035076Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3035192Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3035362Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3035975Z   triton_flex_attention_1708 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3036580Z   triton_flex_attention_1706 0.0107 ms 93.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3037179Z   triton_flex_attention_1709 0.0110 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3037783Z   triton_flex_attention_1704 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3038402Z   triton_flex_attention_1707 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3039001Z   triton_flex_attention_1705 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3039621Z   triton_flex_attention_1724 0.0146 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3040220Z   triton_flex_attention_1716 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3040860Z   triton_flex_attention_1722 0.0160 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3041460Z   triton_flex_attention_1702 0.0166 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3041588Z SingleProcess AUTOTUNE benchmarking takes 0.1988 seconds and 0.5275 seconds precompiling for 24 choices
2025-12-04T09:45:17.3041630Z Autotune Choices Stats:
2025-12-04T09:45:17.3042391Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1743", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.3042624Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3042803Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3043076Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3043704Z   triton_flex_attention_backward_1743 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3044353Z   triton_flex_attention_backward_1737 0.0208 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3044978Z   triton_flex_attention_backward_1734 0.0213 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3045598Z   triton_flex_attention_backward_1735 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3046222Z   triton_flex_attention_backward_1745 0.0232 ms 76.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3046867Z   triton_flex_attention_backward_1744 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3047486Z   triton_flex_attention_backward_1742 0.0249 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3048121Z   triton_flex_attention_backward_1747 0.0252 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3048747Z   triton_flex_attention_backward_1738 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3049384Z   triton_flex_attention_backward_1729 0.0264 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3049512Z SingleProcess AUTOTUNE benchmarking takes 0.2428 seconds and 0.7372 seconds precompiling for 22 choices
2025-12-04T09:45:17.3049588Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3049630Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3049671Z unimplemented []
2025-12-04T09:45:17.3049730Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3049831Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3050448Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3050499Z graph_break []
2025-12-04T09:45:17.3050574Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3050613Z Autotune Choices Stats:
2025-12-04T09:45:17.3051370Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1754", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.3051498Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3051612Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3051786Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3052406Z   triton_flex_attention_1754 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3053011Z   triton_flex_attention_1752 0.0110 ms 90.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3053617Z   triton_flex_attention_1755 0.0114 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3054221Z   triton_flex_attention_1753 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3054826Z   triton_flex_attention_1750 0.0125 ms 79.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3055441Z   triton_flex_attention_1751 0.0143 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3056049Z   triton_flex_attention_1770 0.0149 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3056669Z   triton_flex_attention_1762 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3057275Z   triton_flex_attention_1768 0.0163 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3057883Z   triton_flex_attention_1748 0.0170 ms 58.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3058015Z SingleProcess AUTOTUNE benchmarking takes 0.2060 seconds and 0.4503 seconds precompiling for 24 choices
2025-12-04T09:45:17.3058055Z Autotune Choices Stats:
2025-12-04T09:45:17.3058826Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1789", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.3059052Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3059217Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3059508Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3060138Z   triton_flex_attention_backward_1789 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3060826Z   triton_flex_attention_backward_1783 0.0209 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3061447Z   triton_flex_attention_backward_1780 0.0216 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3062071Z   triton_flex_attention_backward_1781 0.0217 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3062695Z   triton_flex_attention_backward_1791 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3063321Z   triton_flex_attention_backward_1790 0.0235 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3063969Z   triton_flex_attention_backward_1788 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3064595Z   triton_flex_attention_backward_1793 0.0255 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3065239Z   triton_flex_attention_backward_1775 0.0264 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3065864Z   triton_flex_attention_backward_1784 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3065993Z SingleProcess AUTOTUNE benchmarking takes 0.2498 seconds and 0.6949 seconds precompiling for 22 choices
2025-12-04T09:45:17.3066068Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3066111Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3066150Z unimplemented []
2025-12-04T09:45:17.3066213Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3066311Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3066882Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3066921Z graph_break []
2025-12-04T09:45:17.3066995Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3067037Z Autotune Choices Stats:
2025-12-04T09:45:17.3067775Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1800", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.3067915Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3068039Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3068200Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3068810Z   triton_flex_attention_1800 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3069433Z   triton_flex_attention_1798 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3070025Z   triton_flex_attention_1801 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3070666Z   triton_flex_attention_1796 0.0121 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3071268Z   triton_flex_attention_1799 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3071874Z   triton_flex_attention_1816 0.0145 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3072506Z   triton_flex_attention_1797 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3073207Z   triton_flex_attention_1808 0.0152 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3073833Z   triton_flex_attention_1814 0.0161 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3074424Z   triton_flex_attention_1806 0.0168 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3074555Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.5450 seconds precompiling for 24 choices
2025-12-04T09:45:17.3074597Z Autotune Choices Stats:
2025-12-04T09:45:17.3075354Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1835", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.3075574Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3075742Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3076028Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3076672Z   triton_flex_attention_backward_1835 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3077296Z   triton_flex_attention_backward_1829 0.0210 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3077935Z   triton_flex_attention_backward_1826 0.0212 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3078559Z   triton_flex_attention_backward_1827 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3079186Z   triton_flex_attention_backward_1837 0.0231 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3079801Z   triton_flex_attention_backward_1836 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3080464Z   triton_flex_attention_backward_1839 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3081113Z   triton_flex_attention_backward_1834 0.0252 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3081735Z   triton_flex_attention_backward_1830 0.0260 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3082382Z   triton_flex_attention_backward_1821 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3082512Z SingleProcess AUTOTUNE benchmarking takes 0.2508 seconds and 0.7770 seconds precompiling for 22 choices
2025-12-04T09:45:17.3082587Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3082630Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3082670Z unimplemented []
2025-12-04T09:45:17.3082731Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3082832Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3083402Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3083440Z graph_break []
2025-12-04T09:45:17.3083515Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3083555Z Autotune Choices Stats:
2025-12-04T09:45:17.3084295Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1846", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.3084431Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3084546Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3084709Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3085328Z   triton_flex_attention_1846 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3085932Z   triton_flex_attention_1844 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3086557Z   triton_flex_attention_1847 0.0114 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3087163Z   triton_flex_attention_1842 0.0122 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3087763Z   triton_flex_attention_1845 0.0124 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3088365Z   triton_flex_attention_1843 0.0144 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3088971Z   triton_flex_attention_1862 0.0146 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3089592Z   triton_flex_attention_1854 0.0154 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3090192Z   triton_flex_attention_1860 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3090838Z   triton_flex_attention_1840 0.0167 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3090969Z SingleProcess AUTOTUNE benchmarking takes 0.2278 seconds and 0.3492 seconds precompiling for 24 choices
2025-12-04T09:45:17.3091010Z Autotune Choices Stats:
2025-12-04T09:45:17.3091772Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1881", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.3091990Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3092154Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3092430Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3093067Z   triton_flex_attention_backward_1881 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3093723Z   triton_flex_attention_backward_1875 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3094344Z   triton_flex_attention_backward_1873 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3094991Z   triton_flex_attention_backward_1872 0.0216 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3095614Z   triton_flex_attention_backward_1882 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3096242Z   triton_flex_attention_backward_1883 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3096864Z   triton_flex_attention_backward_1880 0.0254 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3097498Z   triton_flex_attention_backward_1885 0.0254 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3098143Z   triton_flex_attention_backward_1876 0.0263 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3098763Z   triton_flex_attention_backward_1867 0.0267 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3098903Z SingleProcess AUTOTUNE benchmarking takes 0.2409 seconds and 0.8665 seconds precompiling for 22 choices
2025-12-04T09:45:17.3098977Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3099030Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3099069Z unimplemented []
2025-12-04T09:45:17.3099131Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3099229Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3099809Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 74), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 28), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 12), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3099849Z graph_break []
2025-12-04T09:45:17.3099925Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3099966Z Autotune Choices Stats:
2025-12-04T09:45:17.3100739Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1892", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.3100867Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3100983Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3101143Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3101767Z   triton_flex_attention_1892 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3102379Z   triton_flex_attention_1890 0.0109 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3102975Z   triton_flex_attention_1893 0.0114 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3103598Z   triton_flex_attention_1888 0.0122 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3104195Z   triton_flex_attention_1891 0.0123 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3104799Z   triton_flex_attention_1889 0.0144 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3105415Z   triton_flex_attention_1908 0.0146 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3106029Z   triton_flex_attention_1900 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3106645Z   triton_flex_attention_1906 0.0161 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3107243Z   triton_flex_attention_1886 0.0167 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3107382Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3466 seconds precompiling for 24 choices
2025-12-04T09:45:17.3107423Z Autotune Choices Stats:
2025-12-04T09:45:17.3108196Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1927", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.3108417Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3108583Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3108858Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3109487Z   triton_flex_attention_backward_1927 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3110109Z   triton_flex_attention_backward_1921 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3110785Z   triton_flex_attention_backward_1918 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3111411Z   triton_flex_attention_backward_1919 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3112061Z   triton_flex_attention_backward_1929 0.0231 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3112761Z   triton_flex_attention_backward_1928 0.0233 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3113378Z   triton_flex_attention_backward_1926 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3114003Z   triton_flex_attention_backward_1931 0.0254 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3114638Z   triton_flex_attention_backward_1922 0.0261 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3115268Z   triton_flex_attention_backward_1913 0.0263 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3115396Z SingleProcess AUTOTUNE benchmarking takes 0.2431 seconds and 0.7860 seconds precompiling for 22 choices
2025-12-04T09:45:17.3115473Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3115525Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3115565Z unimplemented []
2025-12-04T09:45:17.3115625Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3115727Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3116303Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3116342Z graph_break []
2025-12-04T09:45:17.3116418Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3116459Z Autotune Choices Stats:
2025-12-04T09:45:17.3117196Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1938", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.3117322Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3117436Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3117601Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3118218Z   triton_flex_attention_1938 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3118826Z   triton_flex_attention_1936 0.0109 ms 93.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3119439Z   triton_flex_attention_1939 0.0116 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3120038Z   triton_flex_attention_1934 0.0122 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3120686Z   triton_flex_attention_1937 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3121287Z   triton_flex_attention_1935 0.0144 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3121889Z   triton_flex_attention_1954 0.0148 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3122498Z   triton_flex_attention_1946 0.0154 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3123109Z   triton_flex_attention_1952 0.0164 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3123725Z   triton_flex_attention_1944 0.0170 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3123855Z SingleProcess AUTOTUNE benchmarking takes 0.2077 seconds and 0.3245 seconds precompiling for 24 choices
2025-12-04T09:45:17.3123895Z Autotune Choices Stats:
2025-12-04T09:45:17.3124681Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1973", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.3124900Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3125066Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3125343Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3125971Z   triton_flex_attention_backward_1973 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3126599Z   triton_flex_attention_backward_1967 0.0211 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3127231Z   triton_flex_attention_backward_1965 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3127860Z   triton_flex_attention_backward_1964 0.0217 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3128484Z   triton_flex_attention_backward_1975 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3129131Z   triton_flex_attention_backward_1974 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3129744Z   triton_flex_attention_backward_1972 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3130367Z   triton_flex_attention_backward_1977 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3131030Z   triton_flex_attention_backward_1968 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3131662Z   triton_flex_attention_backward_1959 0.0266 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3131805Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.8096 seconds precompiling for 22 choices
2025-12-04T09:45:17.3131878Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3131923Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3131962Z unimplemented []
2025-12-04T09:45:17.3132024Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3132123Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3132686Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3132739Z graph_break []
2025-12-04T09:45:17.3132812Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3132865Z Autotune Choices Stats:
2025-12-04T09:45:17.3133604Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1984", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.3133735Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3133849Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3134011Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3134622Z   triton_flex_attention_1984 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3135215Z   triton_flex_attention_1982 0.0109 ms 95.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3135839Z   triton_flex_attention_1985 0.0113 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3136443Z   triton_flex_attention_1980 0.0122 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3137042Z   triton_flex_attention_1983 0.0124 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3137651Z   triton_flex_attention_1981 0.0142 ms 73.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3138259Z   triton_flex_attention_2000 0.0146 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3138862Z   triton_flex_attention_1992 0.0151 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3139462Z   triton_flex_attention_1998 0.0160 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3140071Z   triton_flex_attention_1978 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3140217Z SingleProcess AUTOTUNE benchmarking takes 0.2059 seconds and 0.3341 seconds precompiling for 24 choices
2025-12-04T09:45:17.3140258Z Autotune Choices Stats:
2025-12-04T09:45:17.3141045Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2019", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.3141276Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3141456Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3141732Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3142363Z   triton_flex_attention_backward_2019 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3142982Z   triton_flex_attention_backward_2013 0.0210 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3143605Z   triton_flex_attention_backward_2010 0.0214 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3144240Z   triton_flex_attention_backward_2011 0.0214 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3144878Z   triton_flex_attention_backward_2021 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3145492Z   triton_flex_attention_backward_2020 0.0233 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3146128Z   triton_flex_attention_backward_2018 0.0250 ms 72.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3146759Z   triton_flex_attention_backward_2023 0.0253 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3147384Z   triton_flex_attention_backward_2014 0.0262 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3148006Z   triton_flex_attention_backward_2005 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3148144Z SingleProcess AUTOTUNE benchmarking takes 0.2422 seconds and 0.7502 seconds precompiling for 22 choices
2025-12-04T09:45:17.3148219Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3148263Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3148302Z unimplemented []
2025-12-04T09:45:17.3148363Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3148464Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3149040Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3149079Z graph_break []
2025-12-04T09:45:17.3149154Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3149204Z Autotune Choices Stats:
2025-12-04T09:45:17.3149951Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2030", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.3150077Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3150192Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3150352Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3151003Z   triton_flex_attention_2030 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3151600Z   triton_flex_attention_2028 0.0109 ms 94.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3152195Z   triton_flex_attention_2031 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3152821Z   triton_flex_attention_2026 0.0126 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3153424Z   triton_flex_attention_2029 0.0127 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3154049Z   triton_flex_attention_2027 0.0142 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3154653Z   triton_flex_attention_2046 0.0147 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3155259Z   triton_flex_attention_2038 0.0152 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3155863Z   triton_flex_attention_2044 0.0162 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3156463Z   triton_flex_attention_2024 0.0165 ms 62.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3156605Z SingleProcess AUTOTUNE benchmarking takes 0.2047 seconds and 0.3631 seconds precompiling for 24 choices
2025-12-04T09:45:17.3156647Z Autotune Choices Stats:
2025-12-04T09:45:17.3157410Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2065", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017799999564886093, "best_triton_pos": 0}
2025-12-04T09:45:17.3157627Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3157794Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3158077Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3158716Z   triton_flex_attention_backward_2065 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3159338Z   triton_flex_attention_backward_2059 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3159959Z   triton_flex_attention_backward_2056 0.0213 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3160607Z   triton_flex_attention_backward_2057 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3161262Z   triton_flex_attention_backward_2067 0.0230 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3161884Z   triton_flex_attention_backward_2066 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3162530Z   triton_flex_attention_backward_2064 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3163153Z   triton_flex_attention_backward_2069 0.0252 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3163775Z   triton_flex_attention_backward_2060 0.0260 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3164404Z   triton_flex_attention_backward_2051 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3164534Z SingleProcess AUTOTUNE benchmarking takes 0.2494 seconds and 0.8153 seconds precompiling for 22 choices
2025-12-04T09:45:17.3164610Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3164664Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3164702Z unimplemented []
2025-12-04T09:45:17.3164763Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3164862Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3165439Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3165479Z graph_break []
2025-12-04T09:45:17.3165552Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3165595Z Autotune Choices Stats:
2025-12-04T09:45:17.3166330Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2076", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.3166469Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3166593Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3166756Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3167365Z   triton_flex_attention_2076 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3167965Z   triton_flex_attention_2074 0.0108 ms 94.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3168568Z   triton_flex_attention_2077 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3169171Z   triton_flex_attention_2072 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3169788Z   triton_flex_attention_2075 0.0125 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3170385Z   triton_flex_attention_2073 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3171065Z   triton_flex_attention_2092 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3171663Z   triton_flex_attention_2084 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3172263Z   triton_flex_attention_2090 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3172862Z   triton_flex_attention_2070 0.0167 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3172994Z SingleProcess AUTOTUNE benchmarking takes 0.2086 seconds and 0.3462 seconds precompiling for 24 choices
2025-12-04T09:45:17.3173037Z Autotune Choices Stats:
2025-12-04T09:45:17.3173794Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2111", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017680000513792038, "best_triton_pos": 0}
2025-12-04T09:45:17.3174038Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3174203Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3174482Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3175135Z   triton_flex_attention_backward_2111 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3175756Z   triton_flex_attention_backward_2105 0.0210 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3176379Z   triton_flex_attention_backward_2102 0.0214 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3176996Z   triton_flex_attention_backward_2103 0.0215 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3177619Z   triton_flex_attention_backward_2113 0.0232 ms 76.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3178262Z   triton_flex_attention_backward_2112 0.0234 ms 75.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3178882Z   triton_flex_attention_backward_2110 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3179516Z   triton_flex_attention_backward_2115 0.0253 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3180140Z   triton_flex_attention_backward_2106 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3180797Z   triton_flex_attention_backward_2097 0.0262 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3180925Z SingleProcess AUTOTUNE benchmarking takes 0.2473 seconds and 0.8010 seconds precompiling for 22 choices
2025-12-04T09:45:17.3181020Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:17.3181069Z Traceback (most recent call last):
2025-12-04T09:45:17.3181224Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:17.3181264Z     self.assertTrue(
2025-12-04T09:45:17.3181371Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:17.3181421Z     raise self.failureException(msg)
2025-12-04T09:45:17.3181551Z AssertionError: False is not true : Log file /tmp/tmprsso7tvz/flex_attention_configs.json was not created
2025-12-04T09:45:17.3181570Z 
2025-12-04T09:45:17.3181646Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.3181811Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.3181814Z 
2025-12-04T09:45:17.3181904Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.3181981Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3182026Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3182064Z unimplemented []
2025-12-04T09:45:17.3182127Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3182715Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:17.3182816Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3182865Z graph_break []
2025-12-04T09:45:17.3182941Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3183431Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:17.3183494Z   current_size = base.storage().size()
2025-12-04T09:45:17.3183537Z Autotune Choices Stats:
2025-12-04T09:45:17.3184280Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.3184410Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3184525Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3184690Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3185304Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3185908Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3186528Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3187125Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3187749Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3188342Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3188943Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3189541Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3190143Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3190785Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3190918Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:17.3190960Z Autotune Choices Stats:
2025-12-04T09:45:17.3191713Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.3191945Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3192120Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3192399Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3193024Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3193636Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3194249Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3194876Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3195510Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3196131Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3196772Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3197399Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3198019Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3198640Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3198781Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:17.3198859Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3198901Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3198941Z unimplemented []
2025-12-04T09:45:17.3199001Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3199112Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3199687Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3199727Z graph_break []
2025-12-04T09:45:17.3199801Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3199853Z Autotune Choices Stats:
2025-12-04T09:45:17.3200642Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.3200769Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3200884Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3201045Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3201656Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3202261Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3202866Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3203491Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3204092Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3204710Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3205309Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3205902Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3206502Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3207102Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3207240Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:17.3207282Z Autotune Choices Stats:
2025-12-04T09:45:17.3208057Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.3208273Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3208440Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3208724Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3209358Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3209980Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3210645Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3211264Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3211909Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3212527Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3213169Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3213789Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3214418Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3215039Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3215170Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:17.3215244Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3215298Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3215336Z unimplemented []
2025-12-04T09:45:17.3215396Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3215496Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3216075Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3216113Z graph_break []
2025-12-04T09:45:17.3216188Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3216230Z Autotune Choices Stats:
2025-12-04T09:45:17.3216967Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.3217105Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3217229Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3217390Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3217997Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3218598Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3219197Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3219796Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3220451Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3221048Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3221673Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3222278Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3222881Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3223482Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3223614Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:17.3223654Z Autotune Choices Stats:
2025-12-04T09:45:17.3224403Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.3224642Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3224806Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3225087Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3225733Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3226358Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3226977Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3227596Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3228221Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3228862Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3229478Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3230116Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3230775Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3231393Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3231523Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:17.3231598Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3231641Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3231679Z unimplemented []
2025-12-04T09:45:17.3231739Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3231840Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3232412Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3232467Z graph_break []
2025-12-04T09:45:17.3232541Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3232584Z Autotune Choices Stats:
2025-12-04T09:45:17.3233326Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:17.3233453Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3233567Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3233739Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3234367Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3234968Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3235569Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3236167Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3236765Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3237385Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3237988Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3238606Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3239208Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3239814Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3239942Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:17.3239984Z Autotune Choices Stats:
2025-12-04T09:45:17.3240772Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.3241002Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3241167Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3241458Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3242078Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3242722Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3243340Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3243960Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3244584Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3245205Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3245848Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3246472Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3247119Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3247738Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3247871Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:17.3247947Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3247991Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3248030Z unimplemented []
2025-12-04T09:45:17.3248093Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3248191Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3248761Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3248800Z graph_break []
2025-12-04T09:45:17.3248874Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3248915Z Autotune Choices Stats:
2025-12-04T09:45:17.3249664Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:17.3249802Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3249916Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3250080Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3250743Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3251371Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3251973Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3252590Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3253187Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3253785Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3254412Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3255010Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3255626Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3256225Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3256356Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:17.3256396Z Autotune Choices Stats:
2025-12-04T09:45:17.3257156Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.3257376Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3257543Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3257832Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3258471Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3259089Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3259722Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3260343Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3260991Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3261615Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3262241Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3262894Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3263515Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3264152Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3264281Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:17.3264356Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3264399Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3264438Z unimplemented []
2025-12-04T09:45:17.3264497Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3264599Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3265169Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3265209Z graph_break []
2025-12-04T09:45:17.3265282Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3265324Z Autotune Choices Stats:
2025-12-04T09:45:17.3266066Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:17.3266204Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3266318Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3266479Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3267102Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3267702Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3268313Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3268912Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3269509Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3270113Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3270753Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3271366Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3271964Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3272588Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3272717Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:17.3272758Z Autotune Choices Stats:
2025-12-04T09:45:17.3273520Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.3273741Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3273910Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3274193Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3274833Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3275479Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3276093Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3276732Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3277355Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3277979Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3278596Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3279220Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3279854Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3280495Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3280635Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:17.3280709Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3280762Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3280803Z unimplemented []
2025-12-04T09:45:17.3280863Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3280963Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3281548Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3281587Z graph_break []
2025-12-04T09:45:17.3281662Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3281703Z Autotune Choices Stats:
2025-12-04T09:45:17.3282454Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:17.3282584Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3282699Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3282859Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3283488Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3284107Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3284709Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3285327Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3285924Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3286524Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3287132Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3287743Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3288353Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3288949Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3289085Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:17.3289125Z Autotune Choices Stats:
2025-12-04T09:45:17.3289886Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.3290105Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3290272Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3290670Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3291300Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3291919Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3292567Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3293185Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3293837Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3294459Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3295962Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3296588Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3297226Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3297850Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3297981Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:17.3298056Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3298101Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3298139Z unimplemented []
2025-12-04T09:45:17.3298201Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3298300Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3298894Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3298933Z graph_break []
2025-12-04T09:45:17.3299009Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3299051Z Autotune Choices Stats:
2025-12-04T09:45:17.3299776Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:17.3299929Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3300043Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3300204Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3300844Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3302199Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3302820Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3303423Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3304033Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3304632Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3305253Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3305853Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3306464Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3307076Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3307207Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:17.3307247Z Autotune Choices Stats:
2025-12-04T09:45:17.3308015Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.3308232Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3308397Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3308684Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3309313Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3309939Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3310591Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3311222Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3311847Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3312479Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3313097Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3313729Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3314356Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3314980Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3315123Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:17.3315198Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3315243Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3315282Z unimplemented []
2025-12-04T09:45:17.3315342Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3315443Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3316015Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3316055Z graph_break []
2025-12-04T09:45:17.3316127Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3316181Z Autotune Choices Stats:
2025-12-04T09:45:17.3316918Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.3317046Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3317162Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3317333Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3317934Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3318535Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3319154Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3319752Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3320356Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3321008Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3321611Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3322222Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3322830Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3323439Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3323581Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:17.3323624Z Autotune Choices Stats:
2025-12-04T09:45:17.3324387Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.3324604Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3324781Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3325062Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3325696Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3326325Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3326938Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3327566Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3328306Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3328928Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3329556Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3330182Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3330841Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3331458Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3331614Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:17.3331688Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3331731Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3331769Z unimplemented []
2025-12-04T09:45:17.3331830Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3331932Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3332518Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3332556Z graph_break []
2025-12-04T09:45:17.3332632Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3332672Z Autotune Choices Stats:
2025-12-04T09:45:17.3333425Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:17.3333554Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3333668Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3333830Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3334446Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3335066Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3335666Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3336286Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3336882Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3337483Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3338150Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3338758Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3339366Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3339967Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3340106Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:17.3340147Z Autotune Choices Stats:
2025-12-04T09:45:17.3340951Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.3341169Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3341333Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3341609Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3342248Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3342870Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3343491Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3344116Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3344748Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3345380Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3345998Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3346632Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3347257Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3350453Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3350593Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:17.3350675Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3350723Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3350793Z unimplemented []
2025-12-04T09:45:17.3350858Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3350964Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3351537Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3351577Z graph_break []
2025-12-04T09:45:17.3351667Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3351713Z Autotune Choices Stats:
2025-12-04T09:45:17.3352453Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:17.3352583Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3352701Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3352874Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3353481Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3354086Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3354703Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3355295Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3355914Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3356514Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3357127Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3357814Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3358410Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3359018Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3359151Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:17.3359194Z Autotune Choices Stats:
2025-12-04T09:45:17.3359942Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.3360182Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3360352Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3360668Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3361295Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3361946Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3362567Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3363199Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3363845Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3364490Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3365106Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3365745Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3366366Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3366987Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3367130Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:17.3367206Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3367251Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3367289Z unimplemented []
2025-12-04T09:45:17.3367353Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3367456Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3368026Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3368075Z graph_break []
2025-12-04T09:45:17.3368150Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3368191Z Autotune Choices Stats:
2025-12-04T09:45:17.3368940Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.3369071Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3369187Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3369348Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3369968Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3370601Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3371201Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3371817Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3372416Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3373031Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3373631Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3374244Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3374846Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3375442Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3375588Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:17.3375629Z Autotune Choices Stats:
2025-12-04T09:45:17.3376379Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.3376614Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3376780Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3377065Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3377691Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3378322Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3378936Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3379555Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3380189Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3380870Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3381533Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3382159Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3382794Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3383414Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3383546Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:17.3383621Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3383677Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3383717Z unimplemented []
2025-12-04T09:45:17.3383779Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3383880Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3384445Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3384483Z graph_break []
2025-12-04T09:45:17.3384558Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3384600Z Autotune Choices Stats:
2025-12-04T09:45:17.3385338Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.3385477Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3385604Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3385767Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3386376Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3386985Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3387585Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3388200Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3388812Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3389409Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3390031Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3390668Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3391283Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3391882Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3392012Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:17.3392054Z Autotune Choices Stats:
2025-12-04T09:45:17.3392825Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.3393043Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3393210Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3393499Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3394140Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3394769Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3395395Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3396011Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3396640Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3397270Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3397887Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3398531Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3399156Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3399783Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3399913Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:17.3399988Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3400030Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3400069Z unimplemented []
2025-12-04T09:45:17.3400128Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3400229Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3400840Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3400896Z graph_break []
2025-12-04T09:45:17.3400971Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3401012Z Autotune Choices Stats:
2025-12-04T09:45:17.3401755Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.3401895Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3402008Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3402171Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3402793Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3403392Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3404011Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3404611Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3405206Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3405813Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3406416Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3407038Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3407640Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3408240Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3408371Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:17.3408411Z Autotune Choices Stats:
2025-12-04T09:45:17.3409165Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.3409393Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3409557Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3409839Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3410495Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3411138Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3411753Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3412382Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3412996Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3413622Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3414252Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3414876Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3415512Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3416132Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3416261Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:17.3416335Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3416380Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3416426Z unimplemented []
2025-12-04T09:45:17.3416491Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3416591Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3417164Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3417203Z graph_break []
2025-12-04T09:45:17.3417278Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3417320Z Autotune Choices Stats:
2025-12-04T09:45:17.3418066Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:17.3418193Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3418309Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3418472Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3419093Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3419706Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3420305Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3420965Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3421565Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3422178Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3422783Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3423398Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3424010Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3424607Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3424736Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:17.3424778Z Autotune Choices Stats:
2025-12-04T09:45:17.3425542Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:17.3425760Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3425926Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3426210Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3426839Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3427461Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3428097Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3428718Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3429353Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3429984Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3430636Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3431274Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3431907Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3432557Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3432686Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:17.3432761Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3432804Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3432842Z unimplemented []
2025-12-04T09:45:17.3432903Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3433003Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3433594Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3433633Z graph_break []
2025-12-04T09:45:17.3433708Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3433749Z Autotune Choices Stats:
2025-12-04T09:45:17.3434489Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.3434625Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3434740Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3434901Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3435513Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3436122Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3436734Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3437335Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3437948Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3438550Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3439164Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3439768Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3440380Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3441014Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3441144Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:17.3441184Z Autotune Choices Stats:
2025-12-04T09:45:17.3441940Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.3442156Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3442322Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3442601Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3443249Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3443886Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3444515Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3445169Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3445790Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3446421Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3447043Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3447680Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3448302Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3448934Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3449063Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:17.3449146Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3449191Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3449230Z unimplemented []
2025-12-04T09:45:17.3449292Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3449391Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3449970Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3450008Z graph_break []
2025-12-04T09:45:17.3450083Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3450125Z Autotune Choices Stats:
2025-12-04T09:45:17.3450912Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.3451041Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3451156Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3451331Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3451943Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3452559Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3453169Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3453780Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3454379Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3454982Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3455584Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3456198Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3456800Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3457406Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3457545Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:17.3457587Z Autotune Choices Stats:
2025-12-04T09:45:17.3458339Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.3458557Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3458732Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3459007Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3459632Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3460259Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3460914Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3461557Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3462193Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3462816Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3463444Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3464068Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3464700Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3465313Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3465449Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:17.3465524Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3465566Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3465603Z unimplemented []
2025-12-04T09:45:17.3465663Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3465765Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3466345Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3466385Z graph_break []
2025-12-04T09:45:17.3466458Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3466498Z Autotune Choices Stats:
2025-12-04T09:45:17.3467245Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:17.3467372Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3467487Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3467647Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3468275Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3468883Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3469488Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3470096Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3470724Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3471322Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3471940Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3472539Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3473149Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3473747Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3473891Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:17.3473932Z Autotune Choices Stats:
2025-12-04T09:45:17.3474699Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:17.3474916Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3475083Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3475361Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3475999Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3476623Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3477257Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3477880Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3478511Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3479140Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3479757Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3480389Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3481046Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3481681Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3481810Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:17.3481884Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3481928Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3481967Z unimplemented []
2025-12-04T09:45:17.3482040Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3482139Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3482711Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3482749Z graph_break []
2025-12-04T09:45:17.3482836Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3482878Z Autotune Choices Stats:
2025-12-04T09:45:17.3483616Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.3483743Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3483856Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3484029Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3484630Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3485225Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3485838Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3486437Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3487052Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3487647Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3488279Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3488879Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3489481Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3490086Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3490214Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:17.3490254Z Autotune Choices Stats:
2025-12-04T09:45:17.3491059Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:17.3491299Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3491485Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3491769Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3492400Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3493046Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3493665Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3494304Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3494927Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3495571Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3496189Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3496821Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3497440Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3498065Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3498204Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:17.3498279Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3498322Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3498361Z unimplemented []
2025-12-04T09:45:17.3498421Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3498521Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3499088Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3499137Z graph_break []
2025-12-04T09:45:17.3499211Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3499252Z Autotune Choices Stats:
2025-12-04T09:45:17.3499999Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.3500127Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3500241Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3500402Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3501061Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3501663Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3502262Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3502871Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3503460Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3504082Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3504683Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3505292Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3505883Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3506481Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3506625Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:17.3506666Z Autotune Choices Stats:
2025-12-04T09:45:17.3507422Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.3507652Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3507819Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3508107Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3508732Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3509368Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3509986Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3510638Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3511282Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3511906Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3512566Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3513189Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3513834Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3514453Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3514582Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:17.3514658Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3514702Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3514755Z unimplemented []
2025-12-04T09:45:17.3514816Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3514916Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3515488Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3515525Z graph_break []
2025-12-04T09:45:17.3515600Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3515639Z Autotune Choices Stats:
2025-12-04T09:45:17.3516373Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.3516512Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3516635Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3516800Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3517409Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3518018Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3518614Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3519216Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3519822Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3520451Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3521098Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3521698Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3522320Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3522923Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3523054Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:17.3523096Z Autotune Choices Stats:
2025-12-04T09:45:17.3523842Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.3524072Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3524238Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3524532Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3525178Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3525794Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3526430Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3527053Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3527670Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3528301Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3528921Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3529568Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3530191Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3530845Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3530976Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:17.3531052Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3531094Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3531133Z unimplemented []
2025-12-04T09:45:17.3531193Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3531297Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3531891Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3531942Z graph_break []
2025-12-04T09:45:17.3532016Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3532058Z Autotune Choices Stats:
2025-12-04T09:45:17.3532783Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:17.3532921Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3533035Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3533193Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3533810Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3534403Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3535014Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3535614Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3536213Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3536819Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3537418Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3538039Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3538639Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3539247Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3539376Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:17.3539419Z Autotune Choices Stats:
2025-12-04T09:45:17.3540183Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.3540446Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3540611Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3540892Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3541525Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3542176Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3542795Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3543423Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3544050Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3544697Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3545330Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3545952Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3546594Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3547214Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3547344Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:17.3547418Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3547461Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3547509Z unimplemented []
2025-12-04T09:45:17.3547572Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3547673Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3548246Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3548283Z graph_break []
2025-12-04T09:45:17.3548359Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3548399Z Autotune Choices Stats:
2025-12-04T09:45:17.3549149Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.3549277Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3549391Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3549559Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3550176Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3550828Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3551428Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3552038Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3552637Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3553238Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3553856Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3554458Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3555077Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3555678Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3555808Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:17.3555850Z Autotune Choices Stats:
2025-12-04T09:45:17.3556620Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.3556839Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3557003Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3557293Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3557926Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3558552Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3559191Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3559809Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3560501Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3561125Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3561745Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3562385Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3563008Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3563647Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3563777Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:17.3563852Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3563893Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3563932Z unimplemented []
2025-12-04T09:45:17.3563992Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3564092Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3564666Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3564707Z graph_break []
2025-12-04T09:45:17.3564780Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3564822Z Autotune Choices Stats:
2025-12-04T09:45:17.3565559Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.3565697Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3565811Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3565972Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3566585Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3567205Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3567814Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3568413Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3569027Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3569628Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3570240Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3570860Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3571476Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3572092Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3572221Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:17.3572263Z Autotune Choices Stats:
2025-12-04T09:45:17.3573029Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.3573248Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3573414Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3573690Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3574312Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3574966Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3575597Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3576224Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3576850Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3577480Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3578101Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3578735Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3579360Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3579991Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3580119Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:17.3580201Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3580247Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3580285Z unimplemented []
2025-12-04T09:45:17.3580346Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3580466Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3581034Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3581071Z graph_break []
2025-12-04T09:45:17.3581147Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3581188Z Autotune Choices Stats:
2025-12-04T09:45:17.3581936Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.3582065Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3582181Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3582356Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3582968Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3583571Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3584183Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3584798Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3585396Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3586023Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3586628Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3587238Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3587838Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3588455Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3588595Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:17.3588636Z Autotune Choices Stats:
2025-12-04T09:45:17.3589394Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.3589609Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3589784Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3590059Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3590716Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3591351Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3591973Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3592603Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3593238Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3593860Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3594481Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3595107Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3595743Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3596361Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3596499Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:17.3596576Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3596617Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3596656Z unimplemented []
2025-12-04T09:45:17.3596716Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3596816Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3597401Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3597441Z graph_break []
2025-12-04T09:45:17.3597516Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3597558Z Autotune Choices Stats:
2025-12-04T09:45:17.3598310Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:17.3598438Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3598552Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3598713Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3599326Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3599941Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3600564Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3601201Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3601802Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3602405Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3603020Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3603622Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3604241Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3604843Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3604982Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:17.3605025Z Autotune Choices Stats:
2025-12-04T09:45:17.3605791Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:17.3606009Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3606174Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3606453Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3607083Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3607705Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3608332Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3608960Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3609590Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3610226Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3610883Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3611530Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3612149Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3612783Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3612912Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:17.3612989Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3613032Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3613083Z unimplemented []
2025-12-04T09:45:17.3613143Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3613244Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3613814Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3613851Z graph_break []
2025-12-04T09:45:17.3613938Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3613979Z Autotune Choices Stats:
2025-12-04T09:45:17.3614716Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.3614844Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3614959Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3615132Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3615745Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3616349Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3616960Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3617561Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3618182Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3618784Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3619403Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3620009Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3620636Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3621247Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3621376Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:17.3621417Z Autotune Choices Stats:
2025-12-04T09:45:17.3622179Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.3622423Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3622590Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3622870Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3623502Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3624128Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3624750Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3625383Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3626008Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3626653Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3627277Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3627909Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3628531Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3629151Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3629290Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:17.3629364Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3629407Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3629445Z unimplemented []
2025-12-04T09:45:17.3629507Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3629605Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3630178Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3630228Z graph_break []
2025-12-04T09:45:17.3630302Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3630343Z Autotune Choices Stats:
2025-12-04T09:45:17.3631128Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:17.3631257Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3631372Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3631532Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3632188Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3632798Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3633406Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3634010Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3634623Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3635263Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3635870Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3636479Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3637084Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3637711Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3637850Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:17.3637892Z Autotune Choices Stats:
2025-12-04T09:45:17.3638656Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:17.3638884Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3639050Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3639337Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3639961Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3640642Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3641265Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3641885Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3642530Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3643147Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3643797Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3644419Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3645052Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3645671Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3645798Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:17.3645874Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3645930Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3645970Z unimplemented []
2025-12-04T09:45:17.3646031Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3646131Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3646709Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3646747Z graph_break []
2025-12-04T09:45:17.3646822Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3646863Z Autotune Choices Stats:
2025-12-04T09:45:17.3647603Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.3647742Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3647856Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3648021Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3648655Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3649266Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3649870Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3650507Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3651116Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3651718Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3652357Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3652956Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3653574Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3654166Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3654297Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:17.3654352Z Autotune Choices Stats:
2025-12-04T09:45:17.3655111Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.3655330Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3655496Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3655785Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3656418Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3657042Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3657676Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3658300Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3658924Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3659562Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3660189Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3660888Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3661510Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3662147Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3662277Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:17.3662351Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3662395Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3662433Z unimplemented []
2025-12-04T09:45:17.3662495Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3662594Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3663171Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3663223Z graph_break []
2025-12-04T09:45:17.3663297Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3663338Z Autotune Choices Stats:
2025-12-04T09:45:17.3664076Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:17.3664217Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3664332Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3664494Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3665114Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3665719Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3666350Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3666952Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3667566Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3668167Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3668781Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3669385Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3669986Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3670656Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3670786Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:17.3670828Z Autotune Choices Stats:
2025-12-04T09:45:17.3671590Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:17.3671820Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3671985Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3672265Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3672896Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3673542Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3674164Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3674800Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3675419Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3676117Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3676748Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3677373Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3678005Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3678627Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3678755Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:17.3678839Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3678883Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3678922Z unimplemented []
2025-12-04T09:45:17.3678983Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3679087Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3679657Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3679694Z graph_break []
2025-12-04T09:45:17.3679769Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3679820Z Autotune Choices Stats:
2025-12-04T09:45:17.3680594Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.3680720Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3680836Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3681010Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3681621Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3682238Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3682842Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3683458Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3684059Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3684675Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3685281Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3685891Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3686516Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3687116Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3687248Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:17.3687289Z Autotune Choices Stats:
2025-12-04T09:45:17.3688043Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.3688260Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3688437Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3688717Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3689346Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3689976Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3690635Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3691256Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3691899Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3692526Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3693151Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3693776Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3694407Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3695041Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3695172Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:17.3695246Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3695290Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3695327Z unimplemented []
2025-12-04T09:45:17.3695388Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3695490Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3696067Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3696108Z graph_break []
2025-12-04T09:45:17.3696181Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3696222Z Autotune Choices Stats:
2025-12-04T09:45:17.3696952Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.3697092Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3697206Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3697369Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3697981Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3698605Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3699206Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3699808Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3700423Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3701051Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3701668Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3702270Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3702885Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3703499Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3703630Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:17.3703672Z Autotune Choices Stats:
2025-12-04T09:45:17.3704447Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.3704666Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3704832Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3705109Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3705752Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3706377Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3707004Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3707635Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3708264Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3708903Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3709527Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3710158Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3710826Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3711459Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3711598Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:17.3711676Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3711720Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3711759Z unimplemented []
2025-12-04T09:45:17.3711820Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3711922Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3712495Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3712533Z graph_break []
2025-12-04T09:45:17.3712621Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3712663Z Autotune Choices Stats:
2025-12-04T09:45:17.3713403Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:17.3713529Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3713645Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3713822Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3714431Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3715038Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3715666Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3716265Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3716873Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3717475Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3718083Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3718694Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3719295Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3719918Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3720048Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:17.3720091Z Autotune Choices Stats:
2025-12-04T09:45:17.3720887Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.3721102Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3721282Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3721563Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3722202Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3722826Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3723447Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3724077Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3724709Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3725334Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3725963Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3726590Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3727222Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3727841Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3727981Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:17.3728055Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3728099Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3728137Z unimplemented []
2025-12-04T09:45:17.3728200Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3728313Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3728887Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3728927Z graph_break []
2025-12-04T09:45:17.3729001Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3729045Z Autotune Choices Stats:
2025-12-04T09:45:17.3729793Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:17.3729922Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3730036Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3730199Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3730849Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3731472Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3732076Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3732704Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3733306Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3733928Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3734530Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3735140Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3735749Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3736348Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3736489Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:17.3736532Z Autotune Choices Stats:
2025-12-04T09:45:17.3737301Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.3737521Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3737687Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3737975Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3738610Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3739241Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3739881Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3740538Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3741178Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3741803Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3742438Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3743062Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3743686Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3744322Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3744449Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:17.3744527Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3744580Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3744619Z unimplemented []
2025-12-04T09:45:17.3744679Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3744781Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3745364Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3745404Z graph_break []
2025-12-04T09:45:17.3745480Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3745521Z Autotune Choices Stats:
2025-12-04T09:45:17.3746257Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.3746384Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3746515Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3746676Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3747291Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3747893Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3748508Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3749107Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3749723Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3750324Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3750986Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3751585Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3752186Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3752802Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3752932Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:17.3752975Z Autotune Choices Stats:
2025-12-04T09:45:17.3753747Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:17.3753976Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3754143Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3754426Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3755071Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3755694Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3756315Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3756943Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3757569Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3758206Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3758824Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3759463Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3760088Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3760740Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3760890Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:17.3760964Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3761008Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3761046Z unimplemented []
2025-12-04T09:45:17.3761108Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3761209Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3761788Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3761838Z graph_break []
2025-12-04T09:45:17.3761913Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3761954Z Autotune Choices Stats:
2025-12-04T09:45:17.3762734Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.3762863Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3762977Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3763139Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3763765Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3764368Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3764972Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3765585Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3766198Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3766813Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3767419Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3768036Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3768641Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3769240Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3769381Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:17.3769421Z Autotune Choices Stats:
2025-12-04T09:45:17.3770186Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.3770447Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3770613Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3770904Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3771537Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3772193Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3772819Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3773438Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3774073Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3774696Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3775338Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3775961Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3776597Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3777222Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3777350Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:17.3777438Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3777482Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3777520Z unimplemented []
2025-12-04T09:45:17.3777581Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3777682Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3778258Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3778298Z graph_break []
2025-12-04T09:45:17.3778372Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3778432Z Autotune Choices Stats:
2025-12-04T09:45:17.3779168Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.3779303Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3779422Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3779585Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3780195Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3780839Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3781447Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3782069Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3782669Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3783284Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3783903Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3784504Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3785113Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3785716Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3785847Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:17.3785907Z Autotune Choices Stats:
2025-12-04T09:45:17.3786666Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.3786884Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3787053Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3787341Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3787986Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3788607Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3789240Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3789852Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3790514Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3791157Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3791800Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3792472Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3793101Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3793743Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3793955Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:17.3794030Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3794074Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3794111Z unimplemented []
2025-12-04T09:45:17.3794173Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3794278Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3794874Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3794914Z graph_break []
2025-12-04T09:45:17.3794993Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3795035Z Autotune Choices Stats:
2025-12-04T09:45:17.3795776Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1708", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.3795917Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3796031Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3796194Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3796826Z   triton_flex_attention_1708 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3797427Z   triton_flex_attention_1706 0.0107 ms 93.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3798033Z   triton_flex_attention_1709 0.0110 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3798633Z   triton_flex_attention_1704 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3799240Z   triton_flex_attention_1707 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3799840Z   triton_flex_attention_1705 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3800489Z   triton_flex_attention_1724 0.0146 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3801105Z   triton_flex_attention_1716 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3801707Z   triton_flex_attention_1722 0.0160 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3802319Z   triton_flex_attention_1702 0.0166 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3802450Z SingleProcess AUTOTUNE benchmarking takes 0.1988 seconds and 0.5275 seconds precompiling for 24 choices
2025-12-04T09:45:17.3802491Z Autotune Choices Stats:
2025-12-04T09:45:17.3803254Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1743", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.3803485Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3803648Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3803929Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3804562Z   triton_flex_attention_backward_1743 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3805192Z   triton_flex_attention_backward_1737 0.0208 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3805812Z   triton_flex_attention_backward_1734 0.0213 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3806466Z   triton_flex_attention_backward_1735 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3807090Z   triton_flex_attention_backward_1745 0.0232 ms 76.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3807720Z   triton_flex_attention_backward_1744 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3808341Z   triton_flex_attention_backward_1742 0.0249 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3808976Z   triton_flex_attention_backward_1747 0.0252 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3809610Z   triton_flex_attention_backward_1738 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3810232Z   triton_flex_attention_backward_1729 0.0264 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3810377Z SingleProcess AUTOTUNE benchmarking takes 0.2428 seconds and 0.7372 seconds precompiling for 22 choices
2025-12-04T09:45:17.3810496Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3810539Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3810578Z unimplemented []
2025-12-04T09:45:17.3810639Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3810741Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3811309Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3811365Z graph_break []
2025-12-04T09:45:17.3811439Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3811482Z Autotune Choices Stats:
2025-12-04T09:45:17.3812221Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1754", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.3812347Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3812463Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3812636Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3813264Z   triton_flex_attention_1754 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3813868Z   triton_flex_attention_1752 0.0110 ms 90.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3814488Z   triton_flex_attention_1755 0.0114 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3815091Z   triton_flex_attention_1753 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3815692Z   triton_flex_attention_1750 0.0125 ms 79.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3816305Z   triton_flex_attention_1751 0.0143 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3816907Z   triton_flex_attention_1770 0.0149 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3817519Z   triton_flex_attention_1762 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3818130Z   triton_flex_attention_1768 0.0163 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3818730Z   triton_flex_attention_1748 0.0170 ms 58.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3818868Z SingleProcess AUTOTUNE benchmarking takes 0.2060 seconds and 0.4503 seconds precompiling for 24 choices
2025-12-04T09:45:17.3818912Z Autotune Choices Stats:
2025-12-04T09:45:17.3819670Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1789", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.3819886Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3820061Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3820339Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3821004Z   triton_flex_attention_backward_1789 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3821640Z   triton_flex_attention_backward_1783 0.0209 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3822276Z   triton_flex_attention_backward_1780 0.0216 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3822896Z   triton_flex_attention_backward_1781 0.0217 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3823540Z   triton_flex_attention_backward_1791 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3824166Z   triton_flex_attention_backward_1790 0.0235 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3824796Z   triton_flex_attention_backward_1788 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3825419Z   triton_flex_attention_backward_1793 0.0255 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3826058Z   triton_flex_attention_backward_1775 0.0264 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3826675Z   triton_flex_attention_backward_1784 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3826806Z SingleProcess AUTOTUNE benchmarking takes 0.2498 seconds and 0.6949 seconds precompiling for 22 choices
2025-12-04T09:45:17.3826881Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3826924Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3826963Z unimplemented []
2025-12-04T09:45:17.3827026Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3827127Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3827712Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3827751Z graph_break []
2025-12-04T09:45:17.3827826Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3827866Z Autotune Choices Stats:
2025-12-04T09:45:17.3828609Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1800", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.3828747Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3828860Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3829024Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3829633Z   triton_flex_attention_1800 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3830251Z   triton_flex_attention_1798 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3830872Z   triton_flex_attention_1801 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3831489Z   triton_flex_attention_1796 0.0121 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3832091Z   triton_flex_attention_1799 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3832695Z   triton_flex_attention_1816 0.0145 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3833307Z   triton_flex_attention_1797 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3833910Z   triton_flex_attention_1808 0.0152 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3834535Z   triton_flex_attention_1814 0.0161 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3835135Z   triton_flex_attention_1806 0.0168 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3835264Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.5450 seconds precompiling for 24 choices
2025-12-04T09:45:17.3835305Z Autotune Choices Stats:
2025-12-04T09:45:17.3836071Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1835", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.3836289Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3836454Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3836734Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3837375Z   triton_flex_attention_backward_1835 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3837997Z   triton_flex_attention_backward_1829 0.0210 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3839920Z   triton_flex_attention_backward_1826 0.0212 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3840592Z   triton_flex_attention_backward_1827 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3841229Z   triton_flex_attention_backward_1837 0.0231 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3841853Z   triton_flex_attention_backward_1836 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3842478Z   triton_flex_attention_backward_1839 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3843108Z   triton_flex_attention_backward_1834 0.0252 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3843731Z   triton_flex_attention_backward_1830 0.0260 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3844370Z   triton_flex_attention_backward_1821 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3844502Z SingleProcess AUTOTUNE benchmarking takes 0.2508 seconds and 0.7770 seconds precompiling for 22 choices
2025-12-04T09:45:17.3844579Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3844620Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3844660Z unimplemented []
2025-12-04T09:45:17.3844720Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3844825Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3845399Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3845449Z graph_break []
2025-12-04T09:45:17.3845524Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3845566Z Autotune Choices Stats:
2025-12-04T09:45:17.3846304Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1846", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.3846432Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3846562Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3846722Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3847356Z   triton_flex_attention_1846 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3847959Z   triton_flex_attention_1844 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3850794Z   triton_flex_attention_1847 0.0114 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3851407Z   triton_flex_attention_1842 0.0122 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3852021Z   triton_flex_attention_1845 0.0124 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3852621Z   triton_flex_attention_1843 0.0144 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3853217Z   triton_flex_attention_1862 0.0146 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3853833Z   triton_flex_attention_1854 0.0154 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3854444Z   triton_flex_attention_1860 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3855067Z   triton_flex_attention_1840 0.0167 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3855201Z SingleProcess AUTOTUNE benchmarking takes 0.2278 seconds and 0.3492 seconds precompiling for 24 choices
2025-12-04T09:45:17.3855244Z Autotune Choices Stats:
2025-12-04T09:45:17.3855999Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1881", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.3856231Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3856402Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3856681Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3857317Z   triton_flex_attention_backward_1881 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3857953Z   triton_flex_attention_backward_1875 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3858577Z   triton_flex_attention_backward_1873 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3859210Z   triton_flex_attention_backward_1872 0.0216 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3859829Z   triton_flex_attention_backward_1882 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3860504Z   triton_flex_attention_backward_1883 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3861127Z   triton_flex_attention_backward_1880 0.0254 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3861754Z   triton_flex_attention_backward_1885 0.0254 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3862401Z   triton_flex_attention_backward_1876 0.0263 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3863027Z   triton_flex_attention_backward_1867 0.0267 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3863171Z SingleProcess AUTOTUNE benchmarking takes 0.2409 seconds and 0.8665 seconds precompiling for 22 choices
2025-12-04T09:45:17.3863248Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3863293Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3863333Z unimplemented []
2025-12-04T09:45:17.3863408Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3863510Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3864085Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 74), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 28), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 12), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3864123Z graph_break []
2025-12-04T09:45:17.3864199Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3864239Z Autotune Choices Stats:
2025-12-04T09:45:17.3864993Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1892", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.3865125Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3865240Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3865409Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3866021Z   triton_flex_attention_1892 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3866636Z   triton_flex_attention_1890 0.0109 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3867243Z   triton_flex_attention_1893 0.0114 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3867863Z   triton_flex_attention_1888 0.0122 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3868463Z   triton_flex_attention_1891 0.0123 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3869070Z   triton_flex_attention_1889 0.0144 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3869672Z   triton_flex_attention_1908 0.0146 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3870275Z   triton_flex_attention_1900 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3870913Z   triton_flex_attention_1906 0.0161 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3871514Z   triton_flex_attention_1886 0.0167 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3871666Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3466 seconds precompiling for 24 choices
2025-12-04T09:45:17.3871706Z Autotune Choices Stats:
2025-12-04T09:45:17.3872468Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1927", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.3872687Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3872854Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3873147Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3873777Z   triton_flex_attention_backward_1927 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3874403Z   triton_flex_attention_backward_1921 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3875039Z   triton_flex_attention_backward_1918 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3875667Z   triton_flex_attention_backward_1919 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3876310Z   triton_flex_attention_backward_1929 0.0231 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3876932Z   triton_flex_attention_backward_1928 0.0233 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3877573Z   triton_flex_attention_backward_1926 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3878195Z   triton_flex_attention_backward_1931 0.0254 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3878826Z   triton_flex_attention_backward_1922 0.0261 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3879454Z   triton_flex_attention_backward_1913 0.0263 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3879588Z SingleProcess AUTOTUNE benchmarking takes 0.2431 seconds and 0.7860 seconds precompiling for 22 choices
2025-12-04T09:45:17.3879673Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3879715Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3879755Z unimplemented []
2025-12-04T09:45:17.3879817Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3879917Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3880536Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3880577Z graph_break []
2025-12-04T09:45:17.3880650Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3880692Z Autotune Choices Stats:
2025-12-04T09:45:17.3881435Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1938", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.3881574Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3881691Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3881853Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3882463Z   triton_flex_attention_1938 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3883078Z   triton_flex_attention_1936 0.0109 ms 93.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3883679Z   triton_flex_attention_1939 0.0116 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3884290Z   triton_flex_attention_1934 0.0122 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3884892Z   triton_flex_attention_1937 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3885497Z   triton_flex_attention_1935 0.0144 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3886127Z   triton_flex_attention_1954 0.0148 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3886733Z   triton_flex_attention_1946 0.0154 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3887337Z   triton_flex_attention_1952 0.0164 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3887946Z   triton_flex_attention_1944 0.0170 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3888075Z SingleProcess AUTOTUNE benchmarking takes 0.2077 seconds and 0.3245 seconds precompiling for 24 choices
2025-12-04T09:45:17.3888128Z Autotune Choices Stats:
2025-12-04T09:45:17.3888895Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1973", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.3889114Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3889281Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3889560Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3890202Z   triton_flex_attention_backward_1973 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3890871Z   triton_flex_attention_backward_1967 0.0211 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3891490Z   triton_flex_attention_backward_1965 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3892129Z   triton_flex_attention_backward_1964 0.0217 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3892755Z   triton_flex_attention_backward_1975 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3893409Z   triton_flex_attention_backward_1974 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3894035Z   triton_flex_attention_backward_1972 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3894667Z   triton_flex_attention_backward_1977 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3895293Z   triton_flex_attention_backward_1968 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3895924Z   triton_flex_attention_backward_1959 0.0266 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3896065Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.8096 seconds precompiling for 22 choices
2025-12-04T09:45:17.3896144Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3896186Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3896225Z unimplemented []
2025-12-04T09:45:17.3896286Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3896389Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3896974Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3897011Z graph_break []
2025-12-04T09:45:17.3897086Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3897126Z Autotune Choices Stats:
2025-12-04T09:45:17.3897874Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1984", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.3898002Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3898115Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3898278Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3898895Z   triton_flex_attention_1984 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3899501Z   triton_flex_attention_1982 0.0109 ms 95.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3900110Z   triton_flex_attention_1985 0.0113 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3900745Z   triton_flex_attention_1980 0.0122 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3901357Z   triton_flex_attention_1983 0.0124 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3901967Z   triton_flex_attention_1981 0.0142 ms 73.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3902573Z   triton_flex_attention_2000 0.0146 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3903186Z   triton_flex_attention_1992 0.0151 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3903787Z   triton_flex_attention_1998 0.0160 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3904389Z   triton_flex_attention_1978 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3904519Z SingleProcess AUTOTUNE benchmarking takes 0.2059 seconds and 0.3341 seconds precompiling for 24 choices
2025-12-04T09:45:17.3904559Z Autotune Choices Stats:
2025-12-04T09:45:17.3905322Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2019", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.3905550Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3905725Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3906006Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3906637Z   triton_flex_attention_backward_2019 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3907272Z   triton_flex_attention_backward_2013 0.0210 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3907891Z   triton_flex_attention_backward_2010 0.0214 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3908515Z   triton_flex_attention_backward_2011 0.0214 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3909148Z   triton_flex_attention_backward_2021 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3909787Z   triton_flex_attention_backward_2020 0.0233 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3910452Z   triton_flex_attention_backward_2018 0.0250 ms 72.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3911074Z   triton_flex_attention_backward_2023 0.0253 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3911712Z   triton_flex_attention_backward_2014 0.0262 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3912330Z   triton_flex_attention_backward_2005 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3912471Z SingleProcess AUTOTUNE benchmarking takes 0.2422 seconds and 0.7502 seconds precompiling for 22 choices
2025-12-04T09:45:17.3912545Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3912588Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3912625Z unimplemented []
2025-12-04T09:45:17.3912687Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3912787Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3913368Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3913419Z graph_break []
2025-12-04T09:45:17.3913491Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3913533Z Autotune Choices Stats:
2025-12-04T09:45:17.3914279Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2030", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.3914408Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3914524Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3914688Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3915289Z   triton_flex_attention_2030 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3915899Z   triton_flex_attention_2028 0.0109 ms 94.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3916503Z   triton_flex_attention_2031 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3917106Z   triton_flex_attention_2026 0.0126 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3917704Z   triton_flex_attention_2029 0.0127 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3918315Z   triton_flex_attention_2027 0.0142 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3918926Z   triton_flex_attention_2046 0.0147 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3919525Z   triton_flex_attention_2038 0.0152 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3920141Z   triton_flex_attention_2044 0.0162 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3920780Z   triton_flex_attention_2024 0.0165 ms 62.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3920922Z SingleProcess AUTOTUNE benchmarking takes 0.2047 seconds and 0.3631 seconds precompiling for 24 choices
2025-12-04T09:45:17.3920964Z Autotune Choices Stats:
2025-12-04T09:45:17.3921718Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2065", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017799999564886093, "best_triton_pos": 0}
2025-12-04T09:45:17.3921937Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3922116Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3922394Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3923035Z   triton_flex_attention_backward_2065 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3923658Z   triton_flex_attention_backward_2059 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3924297Z   triton_flex_attention_backward_2056 0.0213 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3924923Z   triton_flex_attention_backward_2057 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3925553Z   triton_flex_attention_backward_2067 0.0230 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3926294Z   triton_flex_attention_backward_2066 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3926926Z   triton_flex_attention_backward_2064 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3927564Z   triton_flex_attention_backward_2069 0.0252 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3928188Z   triton_flex_attention_backward_2060 0.0260 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3928819Z   triton_flex_attention_backward_2051 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3928948Z SingleProcess AUTOTUNE benchmarking takes 0.2494 seconds and 0.8153 seconds precompiling for 22 choices
2025-12-04T09:45:17.3929023Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3929066Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3929105Z unimplemented []
2025-12-04T09:45:17.3929166Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3929279Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3929851Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3929889Z graph_break []
2025-12-04T09:45:17.3929966Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3930006Z Autotune Choices Stats:
2025-12-04T09:45:17.3930785Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2076", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.3930927Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3931041Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3931216Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3931828Z   triton_flex_attention_2076 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3932447Z   triton_flex_attention_2074 0.0108 ms 94.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3933049Z   triton_flex_attention_2077 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3933654Z   triton_flex_attention_2072 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3934266Z   triton_flex_attention_2075 0.0125 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3934862Z   triton_flex_attention_2073 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3935484Z   triton_flex_attention_2092 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3936085Z   triton_flex_attention_2084 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3936689Z   triton_flex_attention_2090 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3937300Z   triton_flex_attention_2070 0.0167 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3937430Z SingleProcess AUTOTUNE benchmarking takes 0.2086 seconds and 0.3462 seconds precompiling for 24 choices
2025-12-04T09:45:17.3937471Z Autotune Choices Stats:
2025-12-04T09:45:17.3938241Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2111", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017680000513792038, "best_triton_pos": 0}
2025-12-04T09:45:17.3938471Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3938635Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3938917Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3939558Z   triton_flex_attention_backward_2111 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3940189Z   triton_flex_attention_backward_2105 0.0210 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3940851Z   triton_flex_attention_backward_2102 0.0214 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3941492Z   triton_flex_attention_backward_2103 0.0215 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3942115Z   triton_flex_attention_backward_2113 0.0232 ms 76.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3942753Z   triton_flex_attention_backward_2112 0.0234 ms 75.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3943376Z   triton_flex_attention_backward_2110 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3944022Z   triton_flex_attention_backward_2115 0.0253 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3944644Z   triton_flex_attention_backward_2106 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3945280Z   triton_flex_attention_backward_2097 0.0262 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3945410Z SingleProcess AUTOTUNE benchmarking takes 0.2473 seconds and 0.8010 seconds precompiling for 22 choices
2025-12-04T09:45:17.3945485Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3945528Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3945565Z unimplemented []
2025-12-04T09:45:17.3945627Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3945728Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3946296Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.3946346Z graph_break []
2025-12-04T09:45:17.3946418Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3946460Z Autotune Choices Stats:
2025-12-04T09:45:17.3947193Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2122", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.3947321Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3947445Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3947607Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3948227Z   triton_flex_attention_2122 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3948828Z   triton_flex_attention_2120 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3949443Z   triton_flex_attention_2123 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3950035Z   triton_flex_attention_2118 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3950665Z   triton_flex_attention_2121 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3951275Z   triton_flex_attention_2119 0.0142 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3951884Z   triton_flex_attention_2138 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3952510Z   triton_flex_attention_2130 0.0151 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3953113Z   triton_flex_attention_2136 0.0163 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3953724Z   triton_flex_attention_2116 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3953854Z SingleProcess AUTOTUNE benchmarking takes 0.2130 seconds and 0.3464 seconds precompiling for 24 choices
2025-12-04T09:45:17.3953896Z Autotune Choices Stats:
2025-12-04T09:45:17.3954656Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2157", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.3954883Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3955050Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3955331Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3955966Z   triton_flex_attention_backward_2157 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3956609Z   triton_flex_attention_backward_2151 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3957225Z   triton_flex_attention_backward_2148 0.0217 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3957855Z   triton_flex_attention_backward_2149 0.0217 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3958481Z   triton_flex_attention_backward_2159 0.0234 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3959102Z   triton_flex_attention_backward_2158 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3959729Z   triton_flex_attention_backward_2156 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3960354Z   triton_flex_attention_backward_2161 0.0256 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3961026Z   triton_flex_attention_backward_2152 0.0261 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3961646Z   triton_flex_attention_backward_2143 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3961775Z SingleProcess AUTOTUNE benchmarking takes 0.2464 seconds and 0.8851 seconds precompiling for 22 choices
2025-12-04T09:45:17.3961867Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:17.3961916Z Traceback (most recent call last):
2025-12-04T09:45:17.3962074Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:17.3962127Z     self.assertTrue(
2025-12-04T09:45:17.3962236Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:17.3962286Z     raise self.failureException(msg)
2025-12-04T09:45:17.3962413Z AssertionError: False is not true : Log file /tmp/tmpo09mhc5r/flex_attention_configs.json was not created
2025-12-04T09:45:17.3962416Z 
2025-12-04T09:45:17.3962494Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.3962660Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.3962663Z 
2025-12-04T09:45:17.3962752Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.3962827Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3962872Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3962925Z unimplemented []
2025-12-04T09:45:17.3962989Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3963560Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:17.3963660Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3963697Z graph_break []
2025-12-04T09:45:17.3963774Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3964261Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:17.3964325Z   current_size = base.storage().size()
2025-12-04T09:45:17.3964367Z Autotune Choices Stats:
2025-12-04T09:45:17.3965118Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.3965249Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3965363Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3965523Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3966143Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3966742Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3967337Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3967941Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3968536Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3969150Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3969756Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3970365Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3971021Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3971621Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3971765Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:17.3971806Z Autotune Choices Stats:
2025-12-04T09:45:17.3972567Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.3972785Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3972966Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3973244Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3973889Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3974512Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3975135Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3975758Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3976390Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3977012Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3977635Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3978265Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3978887Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3979514Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3979642Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:17.3979719Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3979763Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3979802Z unimplemented []
2025-12-04T09:45:17.3979863Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3979975Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3980591Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3980628Z graph_break []
2025-12-04T09:45:17.3980702Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3980742Z Autotune Choices Stats:
2025-12-04T09:45:17.3981497Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.3981636Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3981750Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3981923Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3982534Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3983137Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3983740Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3984338Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3984951Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3985547Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3986173Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3986770Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3987379Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3987976Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3988105Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:17.3988145Z Autotune Choices Stats:
2025-12-04T09:45:17.3988899Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.3989130Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3989299Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3989580Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3990217Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3990872Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3991490Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3992124Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3992744Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3993378Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3993994Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3994638Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.3995260Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3995891Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3996022Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:17.3996096Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.3996140Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.3996178Z unimplemented []
2025-12-04T09:45:17.3996240Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.3996339Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.3996915Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.3996964Z graph_break []
2025-12-04T09:45:17.3997037Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.3997078Z Autotune Choices Stats:
2025-12-04T09:45:17.3997811Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.3997940Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.3998065Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.3998227Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.3998845Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.3999443Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4000055Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4000679Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4001277Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4001896Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4002495Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4003119Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4003719Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4004347Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4004478Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:17.4004519Z Autotune Choices Stats:
2025-12-04T09:45:17.4005281Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.4005511Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4005679Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4005954Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4006583Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4007224Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4007840Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4008464Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4009090Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4009715Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4010349Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4011009Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4011667Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4012291Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4012421Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:17.4012501Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4012548Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4012595Z unimplemented []
2025-12-04T09:45:17.4012659Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4012783Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4013359Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4013404Z graph_break []
2025-12-04T09:45:17.4013485Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4013529Z Autotune Choices Stats:
2025-12-04T09:45:17.4014268Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:17.4014407Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4014522Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4014690Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4015293Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4015916Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4016517Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4017126Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4017717Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4018319Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4018931Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4019532Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4020159Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4020798Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4020928Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:17.4020968Z Autotune Choices Stats:
2025-12-04T09:45:17.4021735Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.4021953Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4022121Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4022412Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4023040Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4023657Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4024302Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4024924Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4025554Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4026176Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4026796Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4027428Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4028051Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4028689Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4028820Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:17.4028895Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4028938Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4028975Z unimplemented []
2025-12-04T09:45:17.4029037Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4029137Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4029712Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4029753Z graph_break []
2025-12-04T09:45:17.4029830Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4029872Z Autotune Choices Stats:
2025-12-04T09:45:17.4030644Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:17.4030790Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4030903Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4031066Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4031682Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4032287Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4032913Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4033511Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4034118Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4034717Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4035311Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4035924Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4036523Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4037140Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4037270Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:17.4037311Z Autotune Choices Stats:
2025-12-04T09:45:17.4038070Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.4038296Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4038461Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4038737Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4039363Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4039991Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4040649Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4041309Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4041931Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4042565Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4043173Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4043798Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4044427Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4045052Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4045191Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:17.4045268Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4045310Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4045360Z unimplemented []
2025-12-04T09:45:17.4045422Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4045523Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4046089Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4046129Z graph_break []
2025-12-04T09:45:17.4046203Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4046245Z Autotune Choices Stats:
2025-12-04T09:45:17.4046992Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:17.4047120Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4047234Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4047393Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4048012Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4048614Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4049227Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4049837Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4050473Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4051095Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4051692Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4052293Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4052908Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4053524Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4053663Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:17.4053705Z Autotune Choices Stats:
2025-12-04T09:45:17.4054478Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.4054694Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4054861Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4055147Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4055778Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4056420Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4057045Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4057662Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4058311Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4058934Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4059563Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4060189Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4060828Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4061464Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4061595Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:17.4061682Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4061726Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4061764Z unimplemented []
2025-12-04T09:45:17.4061826Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4061925Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4062504Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4062543Z graph_break []
2025-12-04T09:45:17.4062618Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4062658Z Autotune Choices Stats:
2025-12-04T09:45:17.4063404Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:17.4063544Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4063658Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4063821Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4064424Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4065031Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4065635Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4066242Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4066858Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4067456Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4068076Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4068669Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4069272Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4069884Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4070025Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:17.4070066Z Autotune Choices Stats:
2025-12-04T09:45:17.4070860Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.4071078Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4071246Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4071526Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4072180Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4072803Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4073432Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4074064Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4074687Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4075334Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4075967Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4076622Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4077255Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4077884Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4078023Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:17.4078098Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4078141Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4078180Z unimplemented []
2025-12-04T09:45:17.4078240Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4078341Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4078922Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4078961Z graph_break []
2025-12-04T09:45:17.4079034Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4079075Z Autotune Choices Stats:
2025-12-04T09:45:17.4079833Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:17.4079962Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4080077Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4080237Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4080922Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4081533Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4082145Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4082746Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4083355Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4083973Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4084575Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4085205Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4085808Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4086418Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4086547Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:17.4086589Z Autotune Choices Stats:
2025-12-04T09:45:17.4087352Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.4087579Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4087756Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4088032Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4088659Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4089289Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4089908Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4090555Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4091197Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4091823Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4092451Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4093075Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4093708Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4094330Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4094469Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:17.4094544Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4094588Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4094625Z unimplemented []
2025-12-04T09:45:17.4094686Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4094786Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4095359Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4095407Z graph_break []
2025-12-04T09:45:17.4095482Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4095522Z Autotune Choices Stats:
2025-12-04T09:45:17.4096263Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.4096391Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4096505Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4096665Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4097278Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4097890Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4098491Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4099096Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4099693Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4100308Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4100967Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4101566Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4102179Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4102780Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4102921Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:17.4102963Z Autotune Choices Stats:
2025-12-04T09:45:17.4103724Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.4103940Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4104117Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4104391Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4105028Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4105644Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4106270Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4106891Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4107527Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4108154Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4108783Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4109421Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4110041Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4110717Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4110849Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:17.4110923Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4110965Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4111003Z unimplemented []
2025-12-04T09:45:17.4111064Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4111163Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4111746Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4111785Z graph_break []
2025-12-04T09:45:17.4111858Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4111900Z Autotune Choices Stats:
2025-12-04T09:45:17.4112633Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:17.4112781Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4112897Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4113071Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4113680Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4114293Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4114892Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4115491Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4116107Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4116705Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4117319Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4117931Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4118533Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4119147Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4119276Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:17.4119318Z Autotune Choices Stats:
2025-12-04T09:45:17.4120073Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.4120303Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4120510Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4120784Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4121430Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4122062Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4122679Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4123307Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4123930Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4124563Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4125181Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4125809Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4126441Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4127060Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4127199Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:17.4127274Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4127318Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4127355Z unimplemented []
2025-12-04T09:45:17.4127415Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4127514Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4128080Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4128132Z graph_break []
2025-12-04T09:45:17.4128206Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4128246Z Autotune Choices Stats:
2025-12-04T09:45:17.4128986Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:17.4129113Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4129237Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4129397Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4130016Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4130666Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4131289Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4131886Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4132483Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4133099Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4133703Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4134338Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4134937Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4135541Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4135682Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:17.4135723Z Autotune Choices Stats:
2025-12-04T09:45:17.4136478Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.4136695Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4136871Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4137144Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4137776Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4138410Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4139038Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4139663Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4140294Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4140964Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4141596Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4142226Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4142870Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4143490Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4143619Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:17.4143694Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4143736Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4143775Z unimplemented []
2025-12-04T09:45:17.4143835Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4143949Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4144519Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4144558Z graph_break []
2025-12-04T09:45:17.4144632Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4144673Z Autotune Choices Stats:
2025-12-04T09:45:17.4145432Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.4145570Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4145684Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4145845Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4146457Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4147087Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4147689Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4148296Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4148896Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4149496Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4150105Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4150762Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4151390Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4151989Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4152118Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:17.4152160Z Autotune Choices Stats:
2025-12-04T09:45:17.4152925Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.4153145Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4153313Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4153587Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4154224Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4154845Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4155489Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4156105Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4156733Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4157360Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4157979Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4158613Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4159235Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4159872Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4160001Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:17.4160078Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4160122Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4160159Z unimplemented []
2025-12-04T09:45:17.4160219Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4160319Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4160953Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4160995Z graph_break []
2025-12-04T09:45:17.4161069Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4161109Z Autotune Choices Stats:
2025-12-04T09:45:17.4161851Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.4161980Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4162111Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4162273Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4162879Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4163471Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4164121Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4164720Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4165328Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4165928Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4166531Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4167144Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4167752Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4168372Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4168504Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:17.4168545Z Autotune Choices Stats:
2025-12-04T09:45:17.4169311Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.4169540Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4169705Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4169983Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4170652Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4171295Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4171915Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4172551Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4173174Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4173800Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4174416Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4175043Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4175678Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4176312Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4176450Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:17.4176524Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4176568Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4176606Z unimplemented []
2025-12-04T09:45:17.4176678Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4176777Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4177344Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4177382Z graph_break []
2025-12-04T09:45:17.4177455Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4177496Z Autotune Choices Stats:
2025-12-04T09:45:17.4178257Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.4178386Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4178500Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4178661Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4179265Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4179881Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4180515Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4181130Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4181729Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4182342Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4182943Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4183537Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4184151Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4184749Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4184889Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:17.4184930Z Autotune Choices Stats:
2025-12-04T09:45:17.4185698Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.4185915Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4186083Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4186364Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4186991Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4187610Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4188240Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4188864Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4189508Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4190130Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4190806Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4191430Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4192054Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4192698Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4192828Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:17.4192923Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4192965Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4193003Z unimplemented []
2025-12-04T09:45:17.4193064Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4193163Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4193753Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4193792Z graph_break []
2025-12-04T09:45:17.4193868Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4193908Z Autotune Choices Stats:
2025-12-04T09:45:17.4194651Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:17.4194804Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4194920Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4195080Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4195688Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4196305Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4196910Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4197518Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4198129Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4198730Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4199342Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4199943Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4200585Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4201205Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4201335Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:17.4201388Z Autotune Choices Stats:
2025-12-04T09:45:17.4202142Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:17.4202358Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4202525Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4202801Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4203447Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4204062Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4204679Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4205309Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4205936Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4206578Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4207211Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4207843Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4208469Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4209089Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4209229Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:17.4209302Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4209345Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4209383Z unimplemented []
2025-12-04T09:45:17.4209444Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4209543Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4210110Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4210157Z graph_break []
2025-12-04T09:45:17.4210231Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4210273Z Autotune Choices Stats:
2025-12-04T09:45:17.4211081Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.4211210Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4211324Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4211483Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4212096Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4212695Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4213309Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4213911Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4214520Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4215130Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4215734Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4216342Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4216943Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4217545Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4217686Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:17.4217731Z Autotune Choices Stats:
2025-12-04T09:45:17.4218483Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.4218709Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4218883Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4219157Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4219783Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4220457Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4221075Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4221696Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4222338Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4222971Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4223597Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4224221Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4224854Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4225474Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4225612Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:17.4225688Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4225730Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4225770Z unimplemented []
2025-12-04T09:45:17.4225829Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4225931Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4226497Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4226537Z graph_break []
2025-12-04T09:45:17.4226627Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4226668Z Autotune Choices Stats:
2025-12-04T09:45:17.4227419Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.4227546Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4227662Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4227824Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4228435Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4229069Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4229673Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4230281Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4230919Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4231531Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4232153Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4232754Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4233362Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4233960Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4234103Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:17.4234146Z Autotune Choices Stats:
2025-12-04T09:45:17.4234899Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.4235115Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4235293Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4235567Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4236208Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4236827Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4237456Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4238078Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4238709Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4239337Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4239965Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4240653Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4241275Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4241907Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4242036Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:17.4242111Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4242155Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4242193Z unimplemented []
2025-12-04T09:45:17.4242255Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4242355Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4242936Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4242977Z graph_break []
2025-12-04T09:45:17.4243049Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4243090Z Autotune Choices Stats:
2025-12-04T09:45:17.4243829Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:17.4243969Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4244082Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4244255Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4244864Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4245463Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4246068Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4246670Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4247268Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4247863Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4248477Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4249094Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4249697Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4250307Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4250469Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:17.4250511Z Autotune Choices Stats:
2025-12-04T09:45:17.4251265Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:17.4251496Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4251661Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4251941Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4252580Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4253205Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4253828Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4254455Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4255083Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4255713Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4256330Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4256958Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4257599Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4258220Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4258366Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:17.4258443Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4258486Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4258524Z unimplemented []
2025-12-04T09:45:17.4258584Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4258685Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4259263Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4259314Z graph_break []
2025-12-04T09:45:17.4259390Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4259430Z Autotune Choices Stats:
2025-12-04T09:45:17.4260182Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.4260308Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4260458Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4260641Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4261275Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4261875Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4262515Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4263112Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4263711Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4264322Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4264924Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4265535Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4266146Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4266748Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4266887Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:17.4266929Z Autotune Choices Stats:
2025-12-04T09:45:17.4267684Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:17.4267900Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4268079Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4268355Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4268971Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4269600Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4270234Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4270886Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4271532Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4272156Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4272804Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4273424Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4274087Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4274706Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4274838Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:17.4274911Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4274955Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4274992Z unimplemented []
2025-12-04T09:45:17.4275054Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4275153Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4275734Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4275772Z graph_break []
2025-12-04T09:45:17.4275846Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4275886Z Autotune Choices Stats:
2025-12-04T09:45:17.4276621Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.4276760Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4276874Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4277035Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4277639Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4278257Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4278849Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4279461Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4280063Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4280701Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4281317Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4281924Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4282551Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4283150Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4283281Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:17.4283322Z Autotune Choices Stats:
2025-12-04T09:45:17.4284093Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.4284313Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4284478Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4284755Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4285392Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4286014Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4286648Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4287296Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4287919Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4288549Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4289168Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4289797Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4290555Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4291214Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4291344Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:17.4291422Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4291463Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4291502Z unimplemented []
2025-12-04T09:45:17.4291562Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4291662Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4292240Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4292292Z graph_break []
2025-12-04T09:45:17.4292367Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4292409Z Autotune Choices Stats:
2025-12-04T09:45:17.4293141Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.4293267Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4293395Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4293556Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4294157Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4294758Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4295383Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4295985Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4296595Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4297198Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4297802Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4298418Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4299021Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4299642Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4299771Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:17.4299814Z Autotune Choices Stats:
2025-12-04T09:45:17.4300636Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.4300872Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4301041Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4301320Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4301970Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4302600Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4303219Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4303859Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4304483Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4305130Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4305748Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4306362Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4307002Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4307625Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4307768Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:17.4307844Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4307886Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4307923Z unimplemented []
2025-12-04T09:45:17.4307985Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4308096Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4308668Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4308705Z graph_break []
2025-12-04T09:45:17.4308780Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4308820Z Autotune Choices Stats:
2025-12-04T09:45:17.4309567Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:17.4309695Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4309809Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4309971Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4310639Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4311263Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4311863Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4312486Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4313088Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4313699Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4314303Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4314907Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4315518Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4316117Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4316257Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:17.4316297Z Autotune Choices Stats:
2025-12-04T09:45:17.4317078Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.4317295Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4317460Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4317743Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4318378Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4319008Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4319635Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4320252Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4320941Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4321561Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4322205Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4322830Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4323456Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4324096Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4324224Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:17.4324323Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4324365Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4324405Z unimplemented []
2025-12-04T09:45:17.4324465Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4324566Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4325137Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4325178Z graph_break []
2025-12-04T09:45:17.4325250Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4325293Z Autotune Choices Stats:
2025-12-04T09:45:17.4326036Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.4326162Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4326287Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4326448Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4327060Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4327653Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4328265Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4328868Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4329486Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4330086Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4330757Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4331358Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4331960Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4332576Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4332706Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:17.4332761Z Autotune Choices Stats:
2025-12-04T09:45:17.4333516Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.4333743Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4333910Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4334184Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4334821Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4335444Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4336064Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4336687Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4337323Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4337981Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4338600Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4339240Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4339863Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4341448Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4341598Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:17.4341675Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4341718Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4341758Z unimplemented []
2025-12-04T09:45:17.4341819Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4341920Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4342517Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4342572Z graph_break []
2025-12-04T09:45:17.4342647Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4342688Z Autotune Choices Stats:
2025-12-04T09:45:17.4343446Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.4343575Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4343690Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4343851Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4344459Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4345066Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4345704Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4346304Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4346920Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4347533Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4348138Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4348739Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4349346Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4349979Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4350120Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:17.4350161Z Autotune Choices Stats:
2025-12-04T09:45:17.4350952Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.4351182Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4351365Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4351638Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4352267Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4354755Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4355377Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4356026Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4356664Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4357301Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4357932Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4358558Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4359184Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4359808Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4359951Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:17.4360029Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4360075Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4360113Z unimplemented []
2025-12-04T09:45:17.4360177Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4360293Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4360915Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4360955Z graph_break []
2025-12-04T09:45:17.4361047Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4361089Z Autotune Choices Stats:
2025-12-04T09:45:17.4361844Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.4361973Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4362091Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4362253Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4362861Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4363465Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4364072Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4364699Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4365302Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4365914Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4366532Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4367137Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4367739Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4368343Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4368483Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:17.4368526Z Autotune Choices Stats:
2025-12-04T09:45:17.4369301Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.4369519Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4369695Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4369975Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4370661Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4371286Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4371910Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4372534Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4373183Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4373800Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4374439Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4375079Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4375710Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4376333Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4376464Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:17.4376542Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4376585Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4376625Z unimplemented []
2025-12-04T09:45:17.4376687Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4376789Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4377380Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4377421Z graph_break []
2025-12-04T09:45:17.4377496Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4377536Z Autotune Choices Stats:
2025-12-04T09:45:17.4378285Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:17.4378423Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4378538Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4378706Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4379318Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4379924Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4380565Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4381167Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4381786Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4382386Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4383005Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4383618Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4384220Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4384825Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4384954Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:17.4384996Z Autotune Choices Stats:
2025-12-04T09:45:17.4385751Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:17.4385990Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4386157Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4386433Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4387081Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4387727Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4388349Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4388971Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4389590Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4390226Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4390892Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4391532Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4392163Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4392786Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4392916Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:17.4392991Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4393035Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4393073Z unimplemented []
2025-12-04T09:45:17.4393135Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4393236Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4393811Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4393863Z graph_break []
2025-12-04T09:45:17.4393937Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4393978Z Autotune Choices Stats:
2025-12-04T09:45:17.4394730Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.4394858Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4394983Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4395143Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4395765Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4396366Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4396960Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4397555Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4398158Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4398769Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4399373Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4400007Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4400660Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4401260Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4401391Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:17.4401433Z Autotune Choices Stats:
2025-12-04T09:45:17.4402187Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.4402406Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4402587Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4402877Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4403507Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4404143Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4404772Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4405397Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4406021Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4406647Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4407296Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4407913Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4408561Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4409182Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4409312Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:17.4409387Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4409429Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4409468Z unimplemented []
2025-12-04T09:45:17.4409530Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4409630Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4410211Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4410250Z graph_break []
2025-12-04T09:45:17.4410324Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4410365Z Autotune Choices Stats:
2025-12-04T09:45:17.4411170Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:17.4411311Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4411426Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4411586Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4412197Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4412824Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4413430Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4414029Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4414622Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4415214Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4415848Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4416448Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4417066Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4417671Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4417800Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:17.4417842Z Autotune Choices Stats:
2025-12-04T09:45:17.4418598Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:17.4418814Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4418981Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4419259Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4419916Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4420556Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4421204Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4421817Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4422445Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4423081Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4423709Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4424355Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4424981Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4425615Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4425744Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:17.4425819Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4425863Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4425901Z unimplemented []
2025-12-04T09:45:17.4425963Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4426064Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4426632Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4426670Z graph_break []
2025-12-04T09:45:17.4426743Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4426783Z Autotune Choices Stats:
2025-12-04T09:45:17.4427515Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.4427652Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4427768Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4427931Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4428559Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4429161Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4429776Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4430380Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4431022Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4431621Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4432217Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4432836Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4433439Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4434055Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4434186Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:17.4434226Z Autotune Choices Stats:
2025-12-04T09:45:17.4434981Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.4435203Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4435369Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4435645Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4436276Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4436930Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4437556Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4438195Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4438822Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4439453Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4440071Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4440735Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4441387Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4442012Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4442154Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:17.4442230Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4442272Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4442323Z unimplemented []
2025-12-04T09:45:17.4442384Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4442486Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4443067Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4443104Z graph_break []
2025-12-04T09:45:17.4443178Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4443219Z Autotune Choices Stats:
2025-12-04T09:45:17.4443955Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:17.4444082Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4444196Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4444356Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4444985Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4445591Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4446206Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4446820Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4447421Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4448031Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4448640Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4449251Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4449862Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4450528Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4450674Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:17.4450716Z Autotune Choices Stats:
2025-12-04T09:45:17.4451497Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:17.4451714Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4451882Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4452167Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4452800Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4453425Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4454073Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4454698Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4455344Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4455969Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4456590Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4457217Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4457849Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4458487Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4458627Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:17.4458701Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4458744Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4458782Z unimplemented []
2025-12-04T09:45:17.4458845Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4458944Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4459536Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4459575Z graph_break []
2025-12-04T09:45:17.4459651Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4459692Z Autotune Choices Stats:
2025-12-04T09:45:17.4460501Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.4460629Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4460744Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4460907Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4461517Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4462151Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4462754Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4463372Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4463994Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4464597Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4465212Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4465818Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4466447Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4467047Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4467194Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:17.4467234Z Autotune Choices Stats:
2025-12-04T09:45:17.4468003Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.4468220Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4468391Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4468672Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4469330Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4469959Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4470643Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4471275Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4471912Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4472546Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4473175Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4473801Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4474444Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4475091Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4475222Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:17.4475296Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4475339Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4475377Z unimplemented []
2025-12-04T09:45:17.4475437Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4475539Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4476124Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4476163Z graph_break []
2025-12-04T09:45:17.4476236Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4476278Z Autotune Choices Stats:
2025-12-04T09:45:17.4477026Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.4477155Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4477271Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4477432Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4478047Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4478650Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4479268Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4479869Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4480525Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4481135Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4481742Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4482353Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4482966Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4483599Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4483728Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:17.4483770Z Autotune Choices Stats:
2025-12-04T09:45:17.4484526Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.4484756Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4484931Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4485206Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4485836Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4486465Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4487092Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4487729Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4488357Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4488995Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4489631Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4490250Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4490925Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4491546Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4491688Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:17.4491762Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4491804Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4491842Z unimplemented []
2025-12-04T09:45:17.4491916Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4492015Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4492586Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4492637Z graph_break []
2025-12-04T09:45:17.4492712Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4492751Z Autotune Choices Stats:
2025-12-04T09:45:17.4493507Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:17.4493636Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4493749Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4493910Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4494521Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4495125Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4495726Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4496344Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4496934Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4497557Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4498164Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4498774Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4499375Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4499981Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4500128Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:17.4500168Z Autotune Choices Stats:
2025-12-04T09:45:17.4500981Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.4501198Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4501377Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4501657Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4502298Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4502919Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4503542Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4504166Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4504812Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4505429Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4506059Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4506695Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4507318Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4507941Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4508068Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:17.4508143Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4508186Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4508223Z unimplemented []
2025-12-04T09:45:17.4508283Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4508394Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4508978Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4509016Z graph_break []
2025-12-04T09:45:17.4509089Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4509132Z Autotune Choices Stats:
2025-12-04T09:45:17.4509870Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:17.4510009Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4510133Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4510292Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4510944Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4511550Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4512153Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4512760Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4513387Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4513986Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4514621Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4515230Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4515835Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4516438Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4516566Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:17.4516607Z Autotune Choices Stats:
2025-12-04T09:45:17.4517376Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.4517605Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4517770Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4518049Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4518700Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4519324Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4519951Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4520604Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4521233Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4521887Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4522509Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4523155Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4523771Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4524400Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4524528Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:17.4524603Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4524646Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4524685Z unimplemented []
2025-12-04T09:45:17.4524746Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4524846Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4525421Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4525471Z graph_break []
2025-12-04T09:45:17.4525545Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4525585Z Autotune Choices Stats:
2025-12-04T09:45:17.4526337Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.4526474Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4526588Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4526750Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4527371Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4527974Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4528577Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4529178Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4529778Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4530396Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4531037Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4531673Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4532274Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4532874Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4533004Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:17.4533044Z Autotune Choices Stats:
2025-12-04T09:45:17.4533799Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:17.4534031Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4534208Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4534482Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4535123Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4535788Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4536405Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4537028Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4537657Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4538296Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4538940Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4539567Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4540236Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4540918Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4541048Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:17.4541150Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4541241Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4541332Z unimplemented []
2025-12-04T09:45:17.4541433Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4541550Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4542124Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4542219Z graph_break []
2025-12-04T09:45:17.4542335Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4542415Z Autotune Choices Stats:
2025-12-04T09:45:17.4543214Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.4543356Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4543472Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4543649Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4544377Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4545019Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4545621Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4546223Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4546826Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4547436Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4548067Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4548674Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4549293Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4549893Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4550023Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:17.4550064Z Autotune Choices Stats:
2025-12-04T09:45:17.4550861Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.4551079Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4551246Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4551537Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4552181Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4552797Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4553451Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4554071Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4554694Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4555317Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4555939Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4556582Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4557221Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4557860Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4557991Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:17.4558066Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4558109Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4558149Z unimplemented []
2025-12-04T09:45:17.4558209Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4558309Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4558880Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4558921Z graph_break []
2025-12-04T09:45:17.4558995Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4559036Z Autotune Choices Stats:
2025-12-04T09:45:17.4559773Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.4559910Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4560025Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4560193Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4560834Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4561457Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4562083Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4562681Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4563291Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4563898Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4564526Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4565142Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4565762Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4566373Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4566502Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:17.4566544Z Autotune Choices Stats:
2025-12-04T09:45:17.4567301Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.4567517Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4567683Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4567959Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4568594Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4569239Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4569861Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4570556Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4571189Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4571815Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4572442Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4573075Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4573710Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4574345Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4574474Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:17.4574558Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4574601Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4574640Z unimplemented []
2025-12-04T09:45:17.4574702Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4574802Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4575371Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4575408Z graph_break []
2025-12-04T09:45:17.4575484Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4575525Z Autotune Choices Stats:
2025-12-04T09:45:17.4576262Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1708", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.4576389Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4576503Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4576665Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4577285Z   triton_flex_attention_1708 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4577882Z   triton_flex_attention_1706 0.0107 ms 93.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4578493Z   triton_flex_attention_1709 0.0110 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4579105Z   triton_flex_attention_1704 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4579704Z   triton_flex_attention_1707 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4580304Z   triton_flex_attention_1705 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4580947Z   triton_flex_attention_1724 0.0146 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4581576Z   triton_flex_attention_1716 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4582178Z   triton_flex_attention_1722 0.0160 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4582790Z   triton_flex_attention_1702 0.0166 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4582919Z SingleProcess AUTOTUNE benchmarking takes 0.1988 seconds and 0.5275 seconds precompiling for 24 choices
2025-12-04T09:45:17.4582975Z Autotune Choices Stats:
2025-12-04T09:45:17.4583732Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1743", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.4583950Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4584116Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4584396Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4585033Z   triton_flex_attention_backward_1743 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4585665Z   triton_flex_attention_backward_1737 0.0208 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4586293Z   triton_flex_attention_backward_1734 0.0213 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4586927Z   triton_flex_attention_backward_1735 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4587583Z   triton_flex_attention_backward_1745 0.0232 ms 76.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4588205Z   triton_flex_attention_backward_1744 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4588828Z   triton_flex_attention_backward_1742 0.0249 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4589459Z   triton_flex_attention_backward_1747 0.0252 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4590102Z   triton_flex_attention_backward_1738 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4590767Z   triton_flex_attention_backward_1729 0.0264 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4590912Z SingleProcess AUTOTUNE benchmarking takes 0.2428 seconds and 0.7372 seconds precompiling for 22 choices
2025-12-04T09:45:17.4590987Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4591030Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4591069Z unimplemented []
2025-12-04T09:45:17.4591130Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4591230Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4591811Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4591852Z graph_break []
2025-12-04T09:45:17.4591924Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4591966Z Autotune Choices Stats:
2025-12-04T09:45:17.4592708Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1754", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.4592834Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4592948Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4593109Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4593725Z   triton_flex_attention_1754 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4594350Z   triton_flex_attention_1752 0.0110 ms 90.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4594949Z   triton_flex_attention_1755 0.0114 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4595567Z   triton_flex_attention_1753 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4596177Z   triton_flex_attention_1750 0.0125 ms 79.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4596777Z   triton_flex_attention_1751 0.0143 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4597381Z   triton_flex_attention_1770 0.0149 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4597985Z   triton_flex_attention_1762 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4598607Z   triton_flex_attention_1768 0.0163 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4599209Z   triton_flex_attention_1748 0.0170 ms 58.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4599346Z SingleProcess AUTOTUNE benchmarking takes 0.2060 seconds and 0.4503 seconds precompiling for 24 choices
2025-12-04T09:45:17.4599387Z Autotune Choices Stats:
2025-12-04T09:45:17.4600142Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1789", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.4600360Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4600551Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4600826Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4601466Z   triton_flex_attention_backward_1789 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4602093Z   triton_flex_attention_backward_1783 0.0209 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4602741Z   triton_flex_attention_backward_1780 0.0216 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4603357Z   triton_flex_attention_backward_1781 0.0217 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4603994Z   triton_flex_attention_backward_1791 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4604633Z   triton_flex_attention_backward_1790 0.0235 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4605255Z   triton_flex_attention_backward_1788 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4605880Z   triton_flex_attention_backward_1793 0.0255 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4606510Z   triton_flex_attention_backward_1775 0.0264 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4607151Z   triton_flex_attention_backward_1784 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4607279Z SingleProcess AUTOTUNE benchmarking takes 0.2498 seconds and 0.6949 seconds precompiling for 22 choices
2025-12-04T09:45:17.4607354Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4607398Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4607437Z unimplemented []
2025-12-04T09:45:17.4607498Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4607607Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4608175Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4608212Z graph_break []
2025-12-04T09:45:17.4608297Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4608338Z Autotune Choices Stats:
2025-12-04T09:45:17.4609079Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1800", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.4609207Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4609321Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4609483Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4610096Z   triton_flex_attention_1800 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4610744Z   triton_flex_attention_1798 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4611384Z   triton_flex_attention_1801 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4611985Z   triton_flex_attention_1796 0.0121 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4612612Z   triton_flex_attention_1799 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4613219Z   triton_flex_attention_1816 0.0145 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4613826Z   triton_flex_attention_1797 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4614428Z   triton_flex_attention_1808 0.0152 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4615027Z   triton_flex_attention_1814 0.0161 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4615649Z   triton_flex_attention_1806 0.0168 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4615779Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.5450 seconds precompiling for 24 choices
2025-12-04T09:45:17.4615820Z Autotune Choices Stats:
2025-12-04T09:45:17.4616574Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1835", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.4616802Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4616975Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4617250Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4617884Z   triton_flex_attention_backward_1835 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4618509Z   triton_flex_attention_backward_1829 0.0210 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4619121Z   triton_flex_attention_backward_1826 0.0212 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4619769Z   triton_flex_attention_backward_1827 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4620395Z   triton_flex_attention_backward_1837 0.0231 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4621081Z   triton_flex_attention_backward_1836 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4621705Z   triton_flex_attention_backward_1839 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4622328Z   triton_flex_attention_backward_1834 0.0252 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4622961Z   triton_flex_attention_backward_1830 0.0260 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4623590Z   triton_flex_attention_backward_1821 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4623735Z SingleProcess AUTOTUNE benchmarking takes 0.2508 seconds and 0.7770 seconds precompiling for 22 choices
2025-12-04T09:45:17.4623810Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4623866Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4623905Z unimplemented []
2025-12-04T09:45:17.4623965Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4624065Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4624633Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4624686Z graph_break []
2025-12-04T09:45:17.4624759Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4624810Z Autotune Choices Stats:
2025-12-04T09:45:17.4625560Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1846", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.4625688Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4625804Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4625963Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4626578Z   triton_flex_attention_1846 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4627194Z   triton_flex_attention_1844 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4627806Z   triton_flex_attention_1847 0.0114 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4628432Z   triton_flex_attention_1842 0.0122 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4629033Z   triton_flex_attention_1845 0.0124 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4629654Z   triton_flex_attention_1843 0.0144 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4630257Z   triton_flex_attention_1862 0.0146 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4630898Z   triton_flex_attention_1854 0.0154 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4631500Z   triton_flex_attention_1860 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4632103Z   triton_flex_attention_1840 0.0167 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4632247Z SingleProcess AUTOTUNE benchmarking takes 0.2278 seconds and 0.3492 seconds precompiling for 24 choices
2025-12-04T09:45:17.4632289Z Autotune Choices Stats:
2025-12-04T09:45:17.4633061Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1881", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.4633294Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4633460Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4633750Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4634382Z   triton_flex_attention_backward_1881 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4635011Z   triton_flex_attention_backward_1875 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4635632Z   triton_flex_attention_backward_1873 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4636250Z   triton_flex_attention_backward_1872 0.0216 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4636886Z   triton_flex_attention_backward_1882 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4637511Z   triton_flex_attention_backward_1883 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4638154Z   triton_flex_attention_backward_1880 0.0254 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4638778Z   triton_flex_attention_backward_1885 0.0254 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4639417Z   triton_flex_attention_backward_1876 0.0263 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4640037Z   triton_flex_attention_backward_1867 0.0267 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4640163Z SingleProcess AUTOTUNE benchmarking takes 0.2409 seconds and 0.8665 seconds precompiling for 22 choices
2025-12-04T09:45:17.4640241Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4640294Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4640332Z unimplemented []
2025-12-04T09:45:17.4640395Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4640523Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4641114Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 74), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 28), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 12), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4641155Z graph_break []
2025-12-04T09:45:17.4641231Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4641272Z Autotune Choices Stats:
2025-12-04T09:45:17.4642018Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1892", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.4642158Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4642290Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4642452Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4643063Z   triton_flex_attention_1892 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4643663Z   triton_flex_attention_1890 0.0109 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4644268Z   triton_flex_attention_1893 0.0114 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4644867Z   triton_flex_attention_1888 0.0122 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4645503Z   triton_flex_attention_1891 0.0123 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4646101Z   triton_flex_attention_1889 0.0144 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4646725Z   triton_flex_attention_1908 0.0146 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4647333Z   triton_flex_attention_1900 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4647933Z   triton_flex_attention_1906 0.0161 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4648538Z   triton_flex_attention_1886 0.0167 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4648668Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3466 seconds precompiling for 24 choices
2025-12-04T09:45:17.4648708Z Autotune Choices Stats:
2025-12-04T09:45:17.4649478Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1927", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.4649704Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4649868Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4650153Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4650828Z   triton_flex_attention_backward_1927 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4651456Z   triton_flex_attention_backward_1921 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4652070Z   triton_flex_attention_backward_1918 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4652693Z   triton_flex_attention_backward_1919 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4653313Z   triton_flex_attention_backward_1929 0.0231 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4653961Z   triton_flex_attention_backward_1928 0.0233 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4654575Z   triton_flex_attention_backward_1926 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4655223Z   triton_flex_attention_backward_1931 0.0254 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4655847Z   triton_flex_attention_backward_1922 0.0261 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4656468Z   triton_flex_attention_backward_1913 0.0263 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4656597Z SingleProcess AUTOTUNE benchmarking takes 0.2431 seconds and 0.7860 seconds precompiling for 22 choices
2025-12-04T09:45:17.4656672Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4656715Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4656753Z unimplemented []
2025-12-04T09:45:17.4656815Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4656913Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4657490Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4657541Z graph_break []
2025-12-04T09:45:17.4657615Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4657666Z Autotune Choices Stats:
2025-12-04T09:45:17.4658405Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1938", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.4658554Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4658670Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4658830Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4659457Z   triton_flex_attention_1938 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4660071Z   triton_flex_attention_1936 0.0109 ms 93.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4660715Z   triton_flex_attention_1939 0.0116 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4661316Z   triton_flex_attention_1934 0.0122 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4661929Z   triton_flex_attention_1937 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4662542Z   triton_flex_attention_1935 0.0144 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4663145Z   triton_flex_attention_1954 0.0148 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4663768Z   triton_flex_attention_1946 0.0154 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4664376Z   triton_flex_attention_1952 0.0164 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4664977Z   triton_flex_attention_1944 0.0170 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4665106Z SingleProcess AUTOTUNE benchmarking takes 0.2077 seconds and 0.3245 seconds precompiling for 24 choices
2025-12-04T09:45:17.4665148Z Autotune Choices Stats:
2025-12-04T09:45:17.4665903Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1973", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.4666129Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4666306Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4666584Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4667212Z   triton_flex_attention_backward_1973 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4667853Z   triton_flex_attention_backward_1967 0.0211 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4668475Z   triton_flex_attention_backward_1965 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4669099Z   triton_flex_attention_backward_1964 0.0217 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4669725Z   triton_flex_attention_backward_1975 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4670348Z   triton_flex_attention_backward_1974 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4671029Z   triton_flex_attention_backward_1972 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4671654Z   triton_flex_attention_backward_1977 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4672302Z   triton_flex_attention_backward_1968 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4672926Z   triton_flex_attention_backward_1959 0.0266 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4673056Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.8096 seconds precompiling for 22 choices
2025-12-04T09:45:17.4673132Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4673175Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4673215Z unimplemented []
2025-12-04T09:45:17.4673275Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4673376Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4673952Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4673989Z graph_break []
2025-12-04T09:45:17.4674065Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4674117Z Autotune Choices Stats:
2025-12-04T09:45:17.4674883Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1984", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.4675012Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4675126Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4675288Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4675911Z   triton_flex_attention_1984 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4676517Z   triton_flex_attention_1982 0.0109 ms 95.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4677126Z   triton_flex_attention_1985 0.0113 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4677728Z   triton_flex_attention_1980 0.0122 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4678330Z   triton_flex_attention_1983 0.0124 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4678946Z   triton_flex_attention_1981 0.0142 ms 73.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4679552Z   triton_flex_attention_2000 0.0146 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4680164Z   triton_flex_attention_1992 0.0151 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4680818Z   triton_flex_attention_1998 0.0160 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4681416Z   triton_flex_attention_1978 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4681546Z SingleProcess AUTOTUNE benchmarking takes 0.2059 seconds and 0.3341 seconds precompiling for 24 choices
2025-12-04T09:45:17.4681587Z Autotune Choices Stats:
2025-12-04T09:45:17.4682345Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2019", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.4682562Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4682726Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4683014Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4683655Z   triton_flex_attention_backward_2019 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4684280Z   triton_flex_attention_backward_2013 0.0210 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4684923Z   triton_flex_attention_backward_2010 0.0214 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4685543Z   triton_flex_attention_backward_2011 0.0214 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4686195Z   triton_flex_attention_backward_2021 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4686824Z   triton_flex_attention_backward_2020 0.0233 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4687454Z   triton_flex_attention_backward_2018 0.0250 ms 72.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4688105Z   triton_flex_attention_backward_2023 0.0253 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4689451Z   triton_flex_attention_backward_2014 0.0262 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4690792Z   triton_flex_attention_backward_2005 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4691580Z SingleProcess AUTOTUNE benchmarking takes 0.2422 seconds and 0.7502 seconds precompiling for 22 choices
2025-12-04T09:45:17.4691854Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4692021Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4692139Z unimplemented []
2025-12-04T09:45:17.4692270Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4692470Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4693172Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4693810Z graph_break []
2025-12-04T09:45:17.4693942Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4694095Z Autotune Choices Stats:
2025-12-04T09:45:17.4694914Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2030", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.4695836Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4696123Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4696436Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4697267Z   triton_flex_attention_2030 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4698523Z   triton_flex_attention_2028 0.0109 ms 94.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4699788Z   triton_flex_attention_2031 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4701171Z   triton_flex_attention_2026 0.0126 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4702399Z   triton_flex_attention_2029 0.0127 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4703645Z   triton_flex_attention_2027 0.0142 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4704913Z   triton_flex_attention_2046 0.0147 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4706133Z   triton_flex_attention_2038 0.0152 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4707384Z   triton_flex_attention_2044 0.0162 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4708642Z   triton_flex_attention_2024 0.0165 ms 62.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4709397Z SingleProcess AUTOTUNE benchmarking takes 0.2047 seconds and 0.3631 seconds precompiling for 24 choices
2025-12-04T09:45:17.4709601Z Autotune Choices Stats:
2025-12-04T09:45:17.4710457Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2065", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017799999564886093, "best_triton_pos": 0}
2025-12-04T09:45:17.4711464Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4711884Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4712365Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4713331Z   triton_flex_attention_backward_2065 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4714611Z   triton_flex_attention_backward_2059 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4715917Z   triton_flex_attention_backward_2056 0.0213 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4717200Z   triton_flex_attention_backward_2057 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4718501Z   triton_flex_attention_backward_2067 0.0230 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4719781Z   triton_flex_attention_backward_2066 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4721078Z   triton_flex_attention_backward_2064 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4722375Z   triton_flex_attention_backward_2069 0.0252 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4723660Z   triton_flex_attention_backward_2060 0.0260 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4724943Z   triton_flex_attention_backward_2051 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4725735Z SingleProcess AUTOTUNE benchmarking takes 0.2494 seconds and 0.8153 seconds precompiling for 22 choices
2025-12-04T09:45:17.4725977Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4726132Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4726243Z unimplemented []
2025-12-04T09:45:17.4726363Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4726562Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4727272Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4727920Z graph_break []
2025-12-04T09:45:17.4728049Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4728201Z Autotune Choices Stats:
2025-12-04T09:45:17.4728996Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2076", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.4729885Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4730160Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4730521Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4731345Z   triton_flex_attention_2076 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4732575Z   triton_flex_attention_2074 0.0108 ms 94.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4733832Z   triton_flex_attention_2077 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4735066Z   triton_flex_attention_2072 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4736326Z   triton_flex_attention_2075 0.0125 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4737547Z   triton_flex_attention_2073 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4738790Z   triton_flex_attention_2092 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4740068Z   triton_flex_attention_2084 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4741323Z   triton_flex_attention_2090 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4742569Z   triton_flex_attention_2070 0.0167 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4743347Z SingleProcess AUTOTUNE benchmarking takes 0.2086 seconds and 0.3462 seconds precompiling for 24 choices
2025-12-04T09:45:17.4743550Z Autotune Choices Stats:
2025-12-04T09:45:17.4744367Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2111", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017680000513792038, "best_triton_pos": 0}
2025-12-04T09:45:17.4745359Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4745774Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4746249Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4747186Z   triton_flex_attention_backward_2111 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4748488Z   triton_flex_attention_backward_2105 0.0210 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4749766Z   triton_flex_attention_backward_2102 0.0214 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4751081Z   triton_flex_attention_backward_2103 0.0215 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4752372Z   triton_flex_attention_backward_2113 0.0232 ms 76.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4753673Z   triton_flex_attention_backward_2112 0.0234 ms 75.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4754945Z   triton_flex_attention_backward_2110 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4756246Z   triton_flex_attention_backward_2115 0.0253 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4757561Z   triton_flex_attention_backward_2106 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4758833Z   triton_flex_attention_backward_2097 0.0262 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4759641Z SingleProcess AUTOTUNE benchmarking takes 0.2473 seconds and 0.8010 seconds precompiling for 22 choices
2025-12-04T09:45:17.4759881Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4760037Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4760146Z unimplemented []
2025-12-04T09:45:17.4760262Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4760487Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4761221Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4761860Z graph_break []
2025-12-04T09:45:17.4761991Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4762144Z Autotune Choices Stats:
2025-12-04T09:45:17.4762936Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2122", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.4763819Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4764098Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4764406Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4765211Z   triton_flex_attention_2122 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4766472Z   triton_flex_attention_2120 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4767709Z   triton_flex_attention_2123 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4768974Z   triton_flex_attention_2118 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4770206Z   triton_flex_attention_2121 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4771475Z   triton_flex_attention_2119 0.0142 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4772714Z   triton_flex_attention_2138 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4773962Z   triton_flex_attention_2130 0.0151 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4775213Z   triton_flex_attention_2136 0.0163 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4776467Z   triton_flex_attention_2116 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4777247Z SingleProcess AUTOTUNE benchmarking takes 0.2130 seconds and 0.3464 seconds precompiling for 24 choices
2025-12-04T09:45:17.4777452Z Autotune Choices Stats:
2025-12-04T09:45:17.4778275Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2157", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.4779279Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4779694Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4780167Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4781145Z   triton_flex_attention_backward_2157 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4782425Z   triton_flex_attention_backward_2151 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4783726Z   triton_flex_attention_backward_2148 0.0217 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4785010Z   triton_flex_attention_backward_2149 0.0217 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4786315Z   triton_flex_attention_backward_2159 0.0234 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4787591Z   triton_flex_attention_backward_2158 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4788863Z   triton_flex_attention_backward_2156 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4790156Z   triton_flex_attention_backward_2161 0.0256 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4791476Z   triton_flex_attention_backward_2152 0.0261 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4792781Z   triton_flex_attention_backward_2143 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4793554Z SingleProcess AUTOTUNE benchmarking takes 0.2464 seconds and 0.8851 seconds precompiling for 22 choices
2025-12-04T09:45:17.4793793Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4793962Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4794072Z unimplemented []
2025-12-04T09:45:17.4794190Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4794384Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4795101Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4795746Z graph_break []
2025-12-04T09:45:17.4795876Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4796028Z Autotune Choices Stats:
2025-12-04T09:45:17.4796831Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2168", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009840000420808792, "best_triton_pos": 0}
2025-12-04T09:45:17.4797727Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4798001Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4798310Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4799120Z   triton_flex_attention_2168 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4800366Z   triton_flex_attention_2166 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4801676Z   triton_flex_attention_2169 0.0114 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4802905Z   triton_flex_attention_2167 0.0124 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4804170Z   triton_flex_attention_2164 0.0124 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4805417Z   triton_flex_attention_2165 0.0145 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4806655Z   triton_flex_attention_2184 0.0146 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4807897Z   triton_flex_attention_2176 0.0150 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4809132Z   triton_flex_attention_2182 0.0160 ms 61.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4810385Z   triton_flex_attention_2174 0.0167 ms 58.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4811190Z SingleProcess AUTOTUNE benchmarking takes 0.2149 seconds and 0.3567 seconds precompiling for 24 choices
2025-12-04T09:45:17.4811394Z Autotune Choices Stats:
2025-12-04T09:45:17.4812205Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2203", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.4813237Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4813652Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4814127Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4815070Z   triton_flex_attention_backward_2203 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4816366Z   triton_flex_attention_backward_2197 0.0210 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4817663Z   triton_flex_attention_backward_2194 0.0213 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4818956Z   triton_flex_attention_backward_2195 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4820231Z   triton_flex_attention_backward_2205 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4821581Z   triton_flex_attention_backward_2204 0.0233 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4822863Z   triton_flex_attention_backward_2202 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4824145Z   triton_flex_attention_backward_2207 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4825428Z   triton_flex_attention_backward_2198 0.0262 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4826710Z   triton_flex_attention_backward_2189 0.0266 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4827525Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.8512 seconds precompiling for 22 choices
2025-12-04T09:45:17.4827783Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:17.4827959Z Traceback (most recent call last):
2025-12-04T09:45:17.4828192Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:17.4828413Z     self.assertTrue(
2025-12-04T09:45:17.4828581Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:17.4828767Z     raise self.failureException(msg)
2025-12-04T09:45:17.4828982Z AssertionError: False is not true : Log file /tmp/tmp35vgyqua/flex_attention_configs.json was not created
2025-12-04T09:45:17.4831878Z 
2025-12-04T09:45:17.4831966Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.4832253Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.4832461Z 
2025-12-04T09:45:17.4832557Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.4832774Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4832944Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4833064Z unimplemented []
2025-12-04T09:45:17.4833198Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4833908Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:17.4834631Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4834806Z graph_break []
2025-12-04T09:45:17.4834936Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4835560Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:17.4836128Z   current_size = base.storage().size()
2025-12-04T09:45:17.4836254Z Autotune Choices Stats:
2025-12-04T09:45:17.4837066Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.4837959Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4838250Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4838562Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4839378Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4840635Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4841893Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4843124Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4844354Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4845581Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4846816Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4848075Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4849321Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4850605Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4851370Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:17.4851576Z Autotune Choices Stats:
2025-12-04T09:45:17.4852395Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.4853400Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4853821Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4854304Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4855232Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4856530Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4857799Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4859092Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4860392Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4861708Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4862969Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4864233Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4865533Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4866803Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4867599Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:17.4867839Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4867996Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4868106Z unimplemented []
2025-12-04T09:45:17.4868225Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4868433Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4869139Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4869777Z graph_break []
2025-12-04T09:45:17.4869908Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4870061Z Autotune Choices Stats:
2025-12-04T09:45:17.4870879Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.4871766Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4872042Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4872348Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4873164Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4874425Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4875657Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4876922Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4878146Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4879385Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4880671Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4881914Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4883184Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4884413Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4885190Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:17.4885394Z Autotune Choices Stats:
2025-12-04T09:45:17.4886213Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.4887206Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4887623Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4888103Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4889052Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4890335Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4891658Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4892922Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4894229Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4895511Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4896804Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4898079Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4899375Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4900724Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4901506Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:17.4901747Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4901918Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4902029Z unimplemented []
2025-12-04T09:45:17.4902148Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4902344Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4903062Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4903702Z graph_break []
2025-12-04T09:45:17.4903830Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4903983Z Autotune Choices Stats:
2025-12-04T09:45:17.4904796Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.4905682Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4905958Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4906267Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4907077Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4908312Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4909570Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4910839Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4912109Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4913344Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4914575Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4915815Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4917051Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4918305Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4919064Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:17.4919268Z Autotune Choices Stats:
2025-12-04T09:45:17.4920076Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.4921135Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4921550Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4922029Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4922970Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4924251Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4925533Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4926838Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4928114Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4929414Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4930716Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4931992Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4933269Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4934548Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4935344Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:17.4935596Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4935751Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4935860Z unimplemented []
2025-12-04T09:45:17.4935977Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4936169Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4936873Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.4937530Z graph_break []
2025-12-04T09:45:17.4937661Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4937815Z Autotune Choices Stats:
2025-12-04T09:45:17.4938627Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:17.4939512Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4939789Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4940095Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4940939Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4942178Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4943415Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4944682Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4945931Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4947186Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4948419Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4949658Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4950917Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4952180Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4952968Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:17.4953172Z Autotune Choices Stats:
2025-12-04T09:45:17.4954003Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.4955024Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4955440Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4955926Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4956867Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4958148Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4959437Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4960730Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4962036Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4963321Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4964625Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4965898Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4967181Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4968443Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4969224Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:17.4969463Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.4969629Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.4969740Z unimplemented []
2025-12-04T09:45:17.4969856Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.4970048Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.4970794Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.4971432Z graph_break []
2025-12-04T09:45:17.4971560Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.4971712Z Autotune Choices Stats:
2025-12-04T09:45:17.4972523Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:17.4973429Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4973703Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4974012Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4974820Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4976053Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4977286Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4978525Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4979759Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4981020Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.4982277Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4983512Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4984753Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4985993Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4986753Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:17.4986978Z Autotune Choices Stats:
2025-12-04T09:45:17.4987817Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.4988811Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.4989225Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.4989712Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.4990710Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4991982Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4993244Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4994512Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4995814Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4997112Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4998384Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.4999689Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5001039Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5002306Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5003087Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:17.5003325Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5003481Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5003591Z unimplemented []
2025-12-04T09:45:17.5003711Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5003905Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5004607Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5005257Z graph_break []
2025-12-04T09:45:17.5005399Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5005551Z Autotune Choices Stats:
2025-12-04T09:45:17.5006342Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:17.5007237Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5007513Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5007819Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5008636Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5009874Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5011151Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5012386Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5013647Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5014897Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5016142Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5017396Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5018630Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5019881Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5020686Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:17.5020893Z Autotune Choices Stats:
2025-12-04T09:45:17.5021707Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.5022707Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5023136Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5023612Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5024555Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5025863Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5027137Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5028407Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5029688Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5031010Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5032308Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5033601Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5034914Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5036307Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5037090Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:17.5037330Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5037487Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5037597Z unimplemented []
2025-12-04T09:45:17.5037714Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5037908Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5038607Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5039255Z graph_break []
2025-12-04T09:45:17.5039382Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5039547Z Autotune Choices Stats:
2025-12-04T09:45:17.5040354Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:17.5041282Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5041556Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5041862Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5042688Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5043929Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5045186Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5046417Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5047637Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5048913Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5050164Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5051444Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5052704Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5053941Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5054702Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:17.5054914Z Autotune Choices Stats:
2025-12-04T09:45:17.5055741Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.5056739Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5057162Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5057653Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5058617Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5059896Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5061237Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5062523Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5063818Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5065118Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5066407Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5067701Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5068999Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5070296Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5071121Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:17.5071373Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5071537Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5071658Z unimplemented []
2025-12-04T09:45:17.5071781Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5071973Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5072692Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5073336Z graph_break []
2025-12-04T09:45:17.5073466Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5073619Z Autotune Choices Stats:
2025-12-04T09:45:17.5074421Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:17.5075332Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5075632Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5075943Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5076794Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5078054Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5079305Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5080584Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5081816Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5083052Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5084318Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5085562Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5086818Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5088072Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5088838Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:17.5089052Z Autotune Choices Stats:
2025-12-04T09:45:17.5089877Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.5090911Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5091338Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5091831Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5092797Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5094087Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5095489Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5096804Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5098082Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5099350Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5100651Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5101945Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5103228Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5104507Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5105303Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:17.5105542Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5105697Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5105809Z unimplemented []
2025-12-04T09:45:17.5105924Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5106116Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5106817Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5106857Z graph_break []
2025-12-04T09:45:17.5106932Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5106974Z Autotune Choices Stats:
2025-12-04T09:45:17.5107710Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.5107838Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5107956Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5108126Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5108756Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5109353Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5109980Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5110615Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5111209Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5111810Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5112409Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5113039Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5113637Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5114253Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5114405Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:17.5114448Z Autotune Choices Stats:
2025-12-04T09:45:17.5115204Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.5115422Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5115589Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5115875Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5116497Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5117132Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5117749Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5118368Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5118999Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5119618Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5120241Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5120901Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5121555Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5122173Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5122316Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:17.5122392Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5122438Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5122475Z unimplemented []
2025-12-04T09:45:17.5122537Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5122637Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5123233Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5123273Z graph_break []
2025-12-04T09:45:17.5123349Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5123389Z Autotune Choices Stats:
2025-12-04T09:45:17.5124132Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:17.5124261Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5124376Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5124537Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5125147Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5125772Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5126371Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5126993Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5127591Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5128194Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5128799Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5129404Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5131506Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5132110Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5132498Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:17.5132540Z Autotune Choices Stats:
2025-12-04T09:45:17.5133302Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.5133520Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5133685Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5133968Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5134596Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5135215Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5135859Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5136480Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5137121Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5137756Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5138377Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5139003Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5139623Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5140269Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5140401Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:17.5140492Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5140536Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5140591Z unimplemented []
2025-12-04T09:45:17.5140652Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5140754Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5141318Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5141356Z graph_break []
2025-12-04T09:45:17.5141442Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5141485Z Autotune Choices Stats:
2025-12-04T09:45:17.5142221Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:17.5142348Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5142463Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5142627Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5143260Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5143851Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5144472Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5145071Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5145691Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5146284Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5146885Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5147497Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5148096Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5148719Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5148848Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:17.5148889Z Autotune Choices Stats:
2025-12-04T09:45:17.5149639Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.5149879Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5150045Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5150322Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5150975Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5151603Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5152230Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5152873Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5153496Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5154149Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5154776Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5155397Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5156017Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5156639Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5156782Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:17.5156870Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5156914Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5156952Z unimplemented []
2025-12-04T09:45:17.5157014Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5157112Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5157687Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5157736Z graph_break []
2025-12-04T09:45:17.5157811Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5157852Z Autotune Choices Stats:
2025-12-04T09:45:17.5158604Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.5158737Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5158851Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5159012Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5159618Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5160221Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5160851Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5161480Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5162077Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5162697Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5163303Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5163911Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5164517Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5165124Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5165273Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:17.5165327Z Autotune Choices Stats:
2025-12-04T09:45:17.5166082Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.5166309Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5166474Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5166762Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5167391Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5168011Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5168625Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5169246Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5169892Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5170545Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5171197Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5171824Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5172449Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5173070Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5173199Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:17.5173274Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5173329Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5173369Z unimplemented []
2025-12-04T09:45:17.5173431Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5173531Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5174110Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5174148Z graph_break []
2025-12-04T09:45:17.5174222Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5174266Z Autotune Choices Stats:
2025-12-04T09:45:17.5175015Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.5175152Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5175277Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5175437Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5176053Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5176649Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5177250Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5177848Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5178468Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5179059Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5179683Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5180285Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5180919Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5181516Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5181644Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:17.5181687Z Autotune Choices Stats:
2025-12-04T09:45:17.5182465Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.5182684Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5182851Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5183139Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5183775Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5184395Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5185015Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5185635Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5186258Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5186902Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5187522Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5188165Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5188779Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5189397Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5189524Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:17.5189601Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5189644Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5189683Z unimplemented []
2025-12-04T09:45:17.5189743Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5189843Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5190436Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5190490Z graph_break []
2025-12-04T09:45:17.5190564Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5190616Z Autotune Choices Stats:
2025-12-04T09:45:17.5191349Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.5191490Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5191605Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5191767Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5192410Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5193007Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5193608Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5194214Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5194822Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5195422Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5196030Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5196655Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5197255Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5197849Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5197978Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:17.5198018Z Autotune Choices Stats:
2025-12-04T09:45:17.5198788Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.5199016Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5199192Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5199467Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5200094Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5200763Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5201382Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5202003Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5202627Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5203250Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5203895Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5204521Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5205164Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5205785Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5205914Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:17.5205987Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5206032Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5206070Z unimplemented []
2025-12-04T09:45:17.5206132Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5206231Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5206799Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5206836Z graph_break []
2025-12-04T09:45:17.5206910Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5206952Z Autotune Choices Stats:
2025-12-04T09:45:17.5207711Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:17.5207840Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5207956Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5208116Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5208731Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5209342Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5209943Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5210557Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5211158Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5211785Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5212390Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5213000Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5213612Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5214212Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5214341Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:17.5214383Z Autotune Choices Stats:
2025-12-04T09:45:17.5215147Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:17.5215364Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5215530Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5215818Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5216458Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5217071Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5217705Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5218325Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5218947Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5219572Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5220190Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5220852Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5221486Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5222115Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5222244Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:17.5222322Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5222363Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5222402Z unimplemented []
2025-12-04T09:45:17.5222463Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5222563Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5223125Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5223163Z graph_break []
2025-12-04T09:45:17.5223238Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5223279Z Autotune Choices Stats:
2025-12-04T09:45:17.5224010Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.5224147Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5224274Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5224435Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5225047Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5225658Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5226267Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5226871Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5227469Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5228072Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5228699Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5229304Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5229916Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5230562Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5230693Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:17.5230733Z Autotune Choices Stats:
2025-12-04T09:45:17.5231490Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.5231706Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5231871Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5232146Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5232774Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5233409Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5234039Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5234686Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5235309Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5235937Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5236560Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5237200Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5237819Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5238439Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5238569Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:17.5238655Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5238700Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5238738Z unimplemented []
2025-12-04T09:45:17.5238800Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5238898Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5239479Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5239517Z graph_break []
2025-12-04T09:45:17.5239591Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5239634Z Autotune Choices Stats:
2025-12-04T09:45:17.5240369Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.5240532Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5240646Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5240818Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5241442Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5242038Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5242657Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5243268Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5243871Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5244467Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5245071Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5245687Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5246288Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5246895Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5247034Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:17.5247077Z Autotune Choices Stats:
2025-12-04T09:45:17.5247829Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.5248045Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5248213Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5248487Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5249117Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5249751Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5250368Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5251010Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5251652Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5252275Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5252891Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5253518Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5254160Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5254778Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5254919Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:17.5254995Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5255039Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5255078Z unimplemented []
2025-12-04T09:45:17.5255138Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5255240Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5255821Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5255862Z graph_break []
2025-12-04T09:45:17.5255937Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5255978Z Autotune Choices Stats:
2025-12-04T09:45:17.5256717Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:17.5256845Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5256960Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5257123Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5257733Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5258348Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5258950Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5259572Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5260170Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5260800Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5261402Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5262000Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5262625Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5263224Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5263365Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:17.5263404Z Autotune Choices Stats:
2025-12-04T09:45:17.5264170Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:17.5264387Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5264555Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5264835Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5265462Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5266082Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5266721Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5267344Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5267976Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5268604Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5269230Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5269855Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5270510Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5271158Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5271288Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:17.5271362Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5271406Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5271445Z unimplemented []
2025-12-04T09:45:17.5271519Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5271618Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5272197Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5272236Z graph_break []
2025-12-04T09:45:17.5272321Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5272365Z Autotune Choices Stats:
2025-12-04T09:45:17.5273091Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.5273220Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5273334Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5273494Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5274104Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5274703Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5275335Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5275933Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5276555Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5277152Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5277760Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5278366Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5278970Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5279584Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5279715Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:17.5279757Z Autotune Choices Stats:
2025-12-04T09:45:17.5280559Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:17.5280792Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5280969Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5281245Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5281877Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5282498Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5283118Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5283758Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5284384Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5285026Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5285646Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5286275Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5286913Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5287532Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5287668Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:17.5287745Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5287798Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5287838Z unimplemented []
2025-12-04T09:45:17.5287898Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5288000Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5288570Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5288619Z graph_break []
2025-12-04T09:45:17.5288693Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5288736Z Autotune Choices Stats:
2025-12-04T09:45:17.5289489Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.5289618Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5289733Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5289894Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5290599Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5291221Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5291833Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5292473Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5293065Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5293693Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5294293Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5294895Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5295494Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5296106Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5296246Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:17.5296287Z Autotune Choices Stats:
2025-12-04T09:45:17.5297057Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.5297282Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5297450Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5297744Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5298375Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5298998Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5299620Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5300243Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5300927Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5301548Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5302197Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5302821Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5303444Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5304063Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5304195Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:17.5304269Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5304313Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5304363Z unimplemented []
2025-12-04T09:45:17.5304424Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5304526Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5305109Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5305146Z graph_break []
2025-12-04T09:45:17.5305222Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5305263Z Autotune Choices Stats:
2025-12-04T09:45:17.5305993Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.5306128Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5306258Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5306422Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5307032Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5307631Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5308231Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5308830Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5309449Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5310045Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5310716Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5311317Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5311915Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5312522Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5312651Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:17.5312693Z Autotune Choices Stats:
2025-12-04T09:45:17.5313450Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.5313679Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5313845Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5314132Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5314763Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5315386Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5315998Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5316695Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5317321Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5317963Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5318579Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5319221Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5319845Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5320505Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5320635Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:17.5320711Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5320754Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5320793Z unimplemented []
2025-12-04T09:45:17.5320853Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5320954Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5321526Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5321579Z graph_break []
2025-12-04T09:45:17.5321652Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5321705Z Autotune Choices Stats:
2025-12-04T09:45:17.5322443Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:17.5322591Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5322706Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5322865Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5323479Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5324074Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5324675Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5325281Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5325880Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5326499Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5327103Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5327716Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5328317Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5328917Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5329045Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:17.5329087Z Autotune Choices Stats:
2025-12-04T09:45:17.5329868Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.5330093Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5330270Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5330580Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5331206Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5331856Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5332473Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5333094Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5333724Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5334364Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5335004Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5335627Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5336266Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5336883Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5337011Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:17.5337085Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5337129Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5337167Z unimplemented []
2025-12-04T09:45:17.5337229Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5337330Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5337925Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5337962Z graph_break []
2025-12-04T09:45:17.5338038Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5338079Z Autotune Choices Stats:
2025-12-04T09:45:17.5338838Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.5338965Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5339079Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5339238Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5339867Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5340535Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5341140Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5341759Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5342361Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5342969Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5343584Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5344186Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5344805Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5345403Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5345532Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:17.5345574Z Autotune Choices Stats:
2025-12-04T09:45:17.5346330Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.5346551Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5346717Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5347004Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5347634Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5348250Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5348892Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5349512Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5350135Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5350795Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5353340Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5353982Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5354609Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5355263Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5355399Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:17.5355479Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5355526Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5355567Z unimplemented []
2025-12-04T09:45:17.5355630Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5355733Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5356312Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5356353Z graph_break []
2025-12-04T09:45:17.5356427Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5356470Z Autotune Choices Stats:
2025-12-04T09:45:17.5357206Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.5357356Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5357483Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5357643Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5358260Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5358873Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5359492Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5360094Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5360726Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5361322Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5361957Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5362560Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5363173Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5363788Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5363919Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:17.5363960Z Autotune Choices Stats:
2025-12-04T09:45:17.5364721Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.5364940Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5365108Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5365387Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5366028Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5366660Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5367295Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5367936Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5368561Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5369185Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5369809Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5370489Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5371120Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5371753Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5371884Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:17.5371978Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5372024Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5372063Z unimplemented []
2025-12-04T09:45:17.5372126Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5372226Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5372807Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5372845Z graph_break []
2025-12-04T09:45:17.5372921Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5372963Z Autotune Choices Stats:
2025-12-04T09:45:17.5373694Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.5373822Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5373938Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5374110Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5374729Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5375333Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5375947Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5376638Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5377240Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5377841Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5378449Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5379069Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5379670Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5380278Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5380456Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:17.5380500Z Autotune Choices Stats:
2025-12-04T09:45:17.5381250Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.5381467Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5381635Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5381917Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5382546Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5383198Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5383821Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5384454Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5385089Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5385719Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5386351Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5386993Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5387824Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5388465Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5388621Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:17.5388703Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5388746Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5388785Z unimplemented []
2025-12-04T09:45:17.5388846Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5388947Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5389546Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5389587Z graph_break []
2025-12-04T09:45:17.5389661Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5389707Z Autotune Choices Stats:
2025-12-04T09:45:17.5390477Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:17.5390605Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5390722Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5390883Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5391500Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5392129Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5392728Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5393362Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5393966Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5394570Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5395182Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5395791Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5396435Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5397030Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5397168Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:17.5397211Z Autotune Choices Stats:
2025-12-04T09:45:17.5397981Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:17.5398197Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5398364Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5398641Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5399271Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5399898Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5400588Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5401206Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5401844Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5402483Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5403105Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5403728Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5404351Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5404993Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5405122Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:17.5405196Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5405242Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5405293Z unimplemented []
2025-12-04T09:45:17.5405354Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5405454Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5406027Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5406064Z graph_break []
2025-12-04T09:45:17.5406154Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5406196Z Autotune Choices Stats:
2025-12-04T09:45:17.5406935Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.5407062Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5407177Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5407336Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5407939Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5408541Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5409171Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5409770Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5410387Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5411006Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5411608Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5412213Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5412805Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5413435Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5413566Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:17.5413607Z Autotune Choices Stats:
2025-12-04T09:45:17.5414362Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.5414605Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5414770Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5415050Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5415685Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5416310Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5416930Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5417572Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5418198Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5418834Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5419452Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5420080Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5420726Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5421346Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5421500Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:17.5421587Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5421629Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5421669Z unimplemented []
2025-12-04T09:45:17.5421729Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5421831Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5422422Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5422481Z graph_break []
2025-12-04T09:45:17.5422555Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5422599Z Autotune Choices Stats:
2025-12-04T09:45:17.5423342Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:17.5423470Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5423585Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5423744Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5424361Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5424965Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5425567Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5426190Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5426793Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5427415Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5428018Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5428621Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5429224Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5429823Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5429962Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:17.5430020Z Autotune Choices Stats:
2025-12-04T09:45:17.5430830Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:17.5431066Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5431232Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5431529Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5432151Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5432778Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5433404Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5434027Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5434675Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5435299Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5435949Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5436568Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5437193Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5437821Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5437951Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:17.5438025Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5438079Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5438117Z unimplemented []
2025-12-04T09:45:17.5438177Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5438277Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5438863Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5438901Z graph_break []
2025-12-04T09:45:17.5438975Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5439017Z Autotune Choices Stats:
2025-12-04T09:45:17.5439767Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.5439906Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5440019Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5440181Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5440838Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5441443Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5442048Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5442659Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5443272Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5443873Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5444506Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5445109Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5445717Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5446327Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5446458Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:17.5446510Z Autotune Choices Stats:
2025-12-04T09:45:17.5447264Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.5447481Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5447645Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5447932Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5448583Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5449205Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5449818Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5450478Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5451103Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5451750Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5452374Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5453025Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5453653Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5454270Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5454403Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:17.5454480Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5454523Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5454561Z unimplemented []
2025-12-04T09:45:17.5454621Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5454721Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5455296Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5455344Z graph_break []
2025-12-04T09:45:17.5455427Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5455469Z Autotune Choices Stats:
2025-12-04T09:45:17.5456205Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:17.5456341Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5456456Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5456615Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5457239Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5457844Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5458442Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5459054Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5459674Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5460277Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5460930Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5461544Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5462145Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5462748Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5462879Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:17.5462919Z Autotune Choices Stats:
2025-12-04T09:45:17.5463680Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:17.5463908Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5464087Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5464364Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5464994Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5465640Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5466264Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5466886Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5467507Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5468145Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5468785Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5469416Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5470045Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5470707Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5470836Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:17.5470911Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5470955Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5470993Z unimplemented []
2025-12-04T09:45:17.5471053Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5471152Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5471730Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5471767Z graph_break []
2025-12-04T09:45:17.5471842Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5471896Z Autotune Choices Stats:
2025-12-04T09:45:17.5472637Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.5472774Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5472889Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5473064Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5473675Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5474293Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5474897Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5475500Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5476099Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5476717Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5477325Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5477935Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5478545Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5479146Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5479277Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:17.5479318Z Autotune Choices Stats:
2025-12-04T09:45:17.5480078Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.5480295Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5480503Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5480801Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5481434Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5482071Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5482703Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5483325Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5483957Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5484571Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5485208Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5485833Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5486463Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5487093Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5487223Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:17.5487297Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5487340Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5487377Z unimplemented []
2025-12-04T09:45:17.5487439Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5487540Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5488114Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5488154Z graph_break []
2025-12-04T09:45:17.5488227Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5488268Z Autotune Choices Stats:
2025-12-04T09:45:17.5489014Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.5489151Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5489275Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5489437Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5490055Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5490719Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5491319Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5491926Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5492528Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5493126Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5493743Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5494337Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5494945Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5495556Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5495684Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:17.5495726Z Autotune Choices Stats:
2025-12-04T09:45:17.5496483Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.5496701Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5496870Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5497142Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5497789Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5498412Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5499042Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5499697Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5500321Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5500977Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5501609Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5502257Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5502878Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5503514Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5503653Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:17.5503731Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5503774Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5503813Z unimplemented []
2025-12-04T09:45:17.5503872Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5503972Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5504539Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5504576Z graph_break []
2025-12-04T09:45:17.5504651Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5504692Z Autotune Choices Stats:
2025-12-04T09:45:17.5505430Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:17.5505560Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5505674Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5505851Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5506468Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5507065Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5507687Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5508290Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5508880Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5509479Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5510102Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5510755Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5511354Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5511981Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5512111Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:17.5512152Z Autotune Choices Stats:
2025-12-04T09:45:17.5512914Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.5513131Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5513297Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5513573Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5514203Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5514847Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5515473Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5516109Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5516745Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5517369Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5517990Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5518613Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5519255Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5519871Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5520010Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:17.5520085Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5520127Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5520164Z unimplemented []
2025-12-04T09:45:17.5520225Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5520334Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5520936Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5520977Z graph_break []
2025-12-04T09:45:17.5521052Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5521093Z Autotune Choices Stats:
2025-12-04T09:45:17.5521829Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:17.5521957Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5522072Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5522232Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5522854Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5523483Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5524081Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5524703Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5525300Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5525902Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5526502Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5527107Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5527734Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5528336Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5528476Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:17.5528518Z Autotune Choices Stats:
2025-12-04T09:45:17.5529281Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.5529501Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5529667Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5529948Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5530607Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5531233Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5531880Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5532500Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5533149Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5533773Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5534391Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5535013Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5535636Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5536272Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5536404Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:17.5536478Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5536530Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5536570Z unimplemented []
2025-12-04T09:45:17.5536630Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5536730Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5537314Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5537352Z graph_break []
2025-12-04T09:45:17.5537426Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5537468Z Autotune Choices Stats:
2025-12-04T09:45:17.5538196Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.5538323Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5538438Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5538602Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5539215Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5539815Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5540470Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5541073Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5541696Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5542295Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5542907Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5543515Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5544115Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5544738Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5544865Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:17.5544907Z Autotune Choices Stats:
2025-12-04T09:45:17.5545666Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:17.5545890Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5546057Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5546335Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5546962Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5547580Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5548208Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5548852Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5549473Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5550118Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5550773Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5551395Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5552020Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5552633Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5552789Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:17.5552864Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5552907Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5552944Z unimplemented []
2025-12-04T09:45:17.5553006Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5553106Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5553676Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5553728Z graph_break []
2025-12-04T09:45:17.5553802Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5553842Z Autotune Choices Stats:
2025-12-04T09:45:17.5554594Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.5554725Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5554839Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5554999Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5555618Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5556229Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5556837Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5557447Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5558055Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5558665Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5559272Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5559886Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5560531Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5561131Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5561287Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:17.5561330Z Autotune Choices Stats:
2025-12-04T09:45:17.5562082Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.5562316Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5562479Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5562789Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5563425Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5564047Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5564671Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5565292Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5565942Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5566570Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5567204Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5567823Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5568447Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5569067Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5569194Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:17.5569278Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5569320Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5569359Z unimplemented []
2025-12-04T09:45:17.5569419Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5569518Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5570094Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5570133Z graph_break []
2025-12-04T09:45:17.5570208Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5570259Z Autotune Choices Stats:
2025-12-04T09:45:17.5571048Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.5571195Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5571312Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5571474Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5572081Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5572683Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5573276Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5573897Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5574497Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5575100Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5575714Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5576321Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5576928Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5577530Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5577657Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:17.5577711Z Autotune Choices Stats:
2025-12-04T09:45:17.5578468Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.5578685Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5578851Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5579150Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5579789Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5580434Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5581056Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5581677Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5582320Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5582955Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5583590Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5584219Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5584841Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5585464Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5585594Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:17.5585669Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5585712Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5585750Z unimplemented []
2025-12-04T09:45:17.5585811Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5585912Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5586508Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5586547Z graph_break []
2025-12-04T09:45:17.5586619Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5586660Z Autotune Choices Stats:
2025-12-04T09:45:17.5587400Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1708", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.5587539Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5587652Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5587814Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5588456Z   triton_flex_attention_1708 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5589056Z   triton_flex_attention_1706 0.0107 ms 93.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5589658Z   triton_flex_attention_1709 0.0110 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5590260Z   triton_flex_attention_1704 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5590917Z   triton_flex_attention_1707 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5591516Z   triton_flex_attention_1705 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5592132Z   triton_flex_attention_1724 0.0146 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5592743Z   triton_flex_attention_1716 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5593346Z   triton_flex_attention_1722 0.0160 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5593944Z   triton_flex_attention_1702 0.0166 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5594074Z SingleProcess AUTOTUNE benchmarking takes 0.1988 seconds and 0.5275 seconds precompiling for 24 choices
2025-12-04T09:45:17.5594115Z Autotune Choices Stats:
2025-12-04T09:45:17.5594880Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1743", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.5595124Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5595287Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5595565Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5596209Z   triton_flex_attention_backward_1743 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5596840Z   triton_flex_attention_backward_1737 0.0208 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5597459Z   triton_flex_attention_backward_1734 0.0213 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5598077Z   triton_flex_attention_backward_1735 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5598710Z   triton_flex_attention_backward_1745 0.0232 ms 76.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5599353Z   triton_flex_attention_backward_1744 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5599971Z   triton_flex_attention_backward_1742 0.0249 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5600638Z   triton_flex_attention_backward_1747 0.0252 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5601279Z   triton_flex_attention_backward_1738 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5601899Z   triton_flex_attention_backward_1729 0.0264 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5602030Z SingleProcess AUTOTUNE benchmarking takes 0.2428 seconds and 0.7372 seconds precompiling for 22 choices
2025-12-04T09:45:17.5602108Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5602151Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5602189Z unimplemented []
2025-12-04T09:45:17.5602249Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5602350Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5602924Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5602973Z graph_break []
2025-12-04T09:45:17.5603051Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5603094Z Autotune Choices Stats:
2025-12-04T09:45:17.5603849Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1754", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.5603975Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5604090Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5604261Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5604884Z   triton_flex_attention_1754 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5605491Z   triton_flex_attention_1752 0.0110 ms 90.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5606096Z   triton_flex_attention_1755 0.0114 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5606695Z   triton_flex_attention_1753 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5607297Z   triton_flex_attention_1750 0.0125 ms 79.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5607913Z   triton_flex_attention_1751 0.0143 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5608517Z   triton_flex_attention_1770 0.0149 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5609126Z   triton_flex_attention_1762 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5609739Z   triton_flex_attention_1768 0.0163 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5610341Z   triton_flex_attention_1748 0.0170 ms 58.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5610496Z SingleProcess AUTOTUNE benchmarking takes 0.2060 seconds and 0.4503 seconds precompiling for 24 choices
2025-12-04T09:45:17.5610538Z Autotune Choices Stats:
2025-12-04T09:45:17.5611293Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1789", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.5611508Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5611698Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5611991Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5612696Z   triton_flex_attention_backward_1789 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5613327Z   triton_flex_attention_backward_1783 0.0209 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5613950Z   triton_flex_attention_backward_1780 0.0216 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5614579Z   triton_flex_attention_backward_1781 0.0217 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5615210Z   triton_flex_attention_backward_1791 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5615827Z   triton_flex_attention_backward_1790 0.0235 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5616469Z   triton_flex_attention_backward_1788 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5617090Z   triton_flex_attention_backward_1793 0.0255 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5617730Z   triton_flex_attention_backward_1775 0.0264 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5618353Z   triton_flex_attention_backward_1784 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5618485Z SingleProcess AUTOTUNE benchmarking takes 0.2498 seconds and 0.6949 seconds precompiling for 22 choices
2025-12-04T09:45:17.5618559Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5618601Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5618638Z unimplemented []
2025-12-04T09:45:17.5618700Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5618799Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5619364Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5619403Z graph_break []
2025-12-04T09:45:17.5619476Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5619518Z Autotune Choices Stats:
2025-12-04T09:45:17.5620258Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1800", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.5620439Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5620553Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5620715Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5621315Z   triton_flex_attention_1800 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5621948Z   triton_flex_attention_1798 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5622550Z   triton_flex_attention_1801 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5623159Z   triton_flex_attention_1796 0.0121 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5623764Z   triton_flex_attention_1799 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5624370Z   triton_flex_attention_1816 0.0145 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5624992Z   triton_flex_attention_1797 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5625595Z   triton_flex_attention_1808 0.0152 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5626222Z   triton_flex_attention_1814 0.0161 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5626819Z   triton_flex_attention_1806 0.0168 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5626949Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.5450 seconds precompiling for 24 choices
2025-12-04T09:45:17.5626990Z Autotune Choices Stats:
2025-12-04T09:45:17.5627745Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1835", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.5627963Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5628128Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5628405Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5629052Z   triton_flex_attention_backward_1835 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5629672Z   triton_flex_attention_backward_1829 0.0210 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5630316Z   triton_flex_attention_backward_1826 0.0212 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5630955Z   triton_flex_attention_backward_1827 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5631575Z   triton_flex_attention_backward_1837 0.0231 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5632199Z   triton_flex_attention_backward_1836 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5632823Z   triton_flex_attention_backward_1839 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5633470Z   triton_flex_attention_backward_1834 0.0252 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5634093Z   triton_flex_attention_backward_1830 0.0260 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5634737Z   triton_flex_attention_backward_1821 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5634865Z SingleProcess AUTOTUNE benchmarking takes 0.2508 seconds and 0.7770 seconds precompiling for 22 choices
2025-12-04T09:45:17.5634942Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5634984Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5635023Z unimplemented []
2025-12-04T09:45:17.5635084Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5635185Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5635755Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5635795Z graph_break []
2025-12-04T09:45:17.5635871Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5635914Z Autotune Choices Stats:
2025-12-04T09:45:17.5636655Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1846", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.5636781Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5636908Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5637068Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5637692Z   triton_flex_attention_1846 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5638288Z   triton_flex_attention_1844 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5638911Z   triton_flex_attention_1847 0.0114 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5639511Z   triton_flex_attention_1842 0.0122 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5640109Z   triton_flex_attention_1845 0.0124 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5640747Z   triton_flex_attention_1843 0.0144 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5641349Z   triton_flex_attention_1862 0.0146 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5641978Z   triton_flex_attention_1854 0.0154 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5642584Z   triton_flex_attention_1860 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5643210Z   triton_flex_attention_1840 0.0167 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5643338Z SingleProcess AUTOTUNE benchmarking takes 0.2278 seconds and 0.3492 seconds precompiling for 24 choices
2025-12-04T09:45:17.5643383Z Autotune Choices Stats:
2025-12-04T09:45:17.5644139Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1881", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.5644354Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5644521Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5644800Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5645432Z   triton_flex_attention_backward_1881 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5646074Z   triton_flex_attention_backward_1875 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5646696Z   triton_flex_attention_backward_1873 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5647339Z   triton_flex_attention_backward_1872 0.0216 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5647962Z   triton_flex_attention_backward_1882 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5648590Z   triton_flex_attention_backward_1883 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5649216Z   triton_flex_attention_backward_1880 0.0254 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5649835Z   triton_flex_attention_backward_1885 0.0254 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5650510Z   triton_flex_attention_backward_1876 0.0263 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5651124Z   triton_flex_attention_backward_1867 0.0267 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5651267Z SingleProcess AUTOTUNE benchmarking takes 0.2409 seconds and 0.8665 seconds precompiling for 22 choices
2025-12-04T09:45:17.5651341Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5651384Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5651421Z unimplemented []
2025-12-04T09:45:17.5651494Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5651594Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5652162Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 74), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 28), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 12), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5652200Z graph_break []
2025-12-04T09:45:17.5652274Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5652315Z Autotune Choices Stats:
2025-12-04T09:45:17.5653056Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1892", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.5653185Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5653299Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5653459Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5654083Z   triton_flex_attention_1892 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5654692Z   triton_flex_attention_1890 0.0109 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5655286Z   triton_flex_attention_1893 0.0114 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5655909Z   triton_flex_attention_1888 0.0122 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5656508Z   triton_flex_attention_1891 0.0123 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5657110Z   triton_flex_attention_1889 0.0144 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5657718Z   triton_flex_attention_1908 0.0146 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5658322Z   triton_flex_attention_1900 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5658956Z   triton_flex_attention_1906 0.0161 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5659558Z   triton_flex_attention_1886 0.0167 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5659696Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3466 seconds precompiling for 24 choices
2025-12-04T09:45:17.5659736Z Autotune Choices Stats:
2025-12-04T09:45:17.5660524Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1927", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.5660743Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5660908Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5661184Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5661807Z   triton_flex_attention_backward_1927 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5662422Z   triton_flex_attention_backward_1921 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5663070Z   triton_flex_attention_backward_1918 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5663690Z   triton_flex_attention_backward_1919 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5664338Z   triton_flex_attention_backward_1929 0.0231 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5664966Z   triton_flex_attention_backward_1928 0.0233 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5665584Z   triton_flex_attention_backward_1926 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5666239Z   triton_flex_attention_backward_1931 0.0254 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5666859Z   triton_flex_attention_backward_1922 0.0261 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5667505Z   triton_flex_attention_backward_1913 0.0263 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5667634Z SingleProcess AUTOTUNE benchmarking takes 0.2431 seconds and 0.7860 seconds precompiling for 22 choices
2025-12-04T09:45:17.5667719Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5667760Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5667798Z unimplemented []
2025-12-04T09:45:17.5667858Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5667959Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5668537Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5668578Z graph_break []
2025-12-04T09:45:17.5668651Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5668693Z Autotune Choices Stats:
2025-12-04T09:45:17.5669436Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1938", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.5669561Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5669677Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5669835Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5670475Z   triton_flex_attention_1938 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5671102Z   triton_flex_attention_1936 0.0109 ms 93.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5671708Z   triton_flex_attention_1939 0.0116 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5672320Z   triton_flex_attention_1934 0.0122 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5672929Z   triton_flex_attention_1937 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5673530Z   triton_flex_attention_1935 0.0144 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5674136Z   triton_flex_attention_1954 0.0148 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5674736Z   triton_flex_attention_1946 0.0154 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5675356Z   triton_flex_attention_1952 0.0164 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5675968Z   triton_flex_attention_1944 0.0170 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5676096Z SingleProcess AUTOTUNE benchmarking takes 0.2077 seconds and 0.3245 seconds precompiling for 24 choices
2025-12-04T09:45:17.5676153Z Autotune Choices Stats:
2025-12-04T09:45:17.5676917Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1973", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.5677132Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5677300Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5677574Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5678207Z   triton_flex_attention_backward_1973 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5678827Z   triton_flex_attention_backward_1967 0.0211 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5679446Z   triton_flex_attention_backward_1965 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5680082Z   triton_flex_attention_backward_1964 0.0217 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5680747Z   triton_flex_attention_backward_1975 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5681394Z   triton_flex_attention_backward_1974 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5682006Z   triton_flex_attention_backward_1972 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5682629Z   triton_flex_attention_backward_1977 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5683251Z   triton_flex_attention_backward_1968 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5683884Z   triton_flex_attention_backward_1959 0.0266 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5684027Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.8096 seconds precompiling for 22 choices
2025-12-04T09:45:17.5684103Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5684147Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5684186Z unimplemented []
2025-12-04T09:45:17.5684249Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5684349Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5684936Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5684974Z graph_break []
2025-12-04T09:45:17.5685049Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5685090Z Autotune Choices Stats:
2025-12-04T09:45:17.5685837Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1984", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.5685967Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5686080Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5686241Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5686852Z   triton_flex_attention_1984 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5687457Z   triton_flex_attention_1982 0.0109 ms 95.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5688077Z   triton_flex_attention_1985 0.0113 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5688677Z   triton_flex_attention_1980 0.0122 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5689287Z   triton_flex_attention_1983 0.0124 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5689904Z   triton_flex_attention_1981 0.0142 ms 73.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5690540Z   triton_flex_attention_2000 0.0146 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5691140Z   triton_flex_attention_1992 0.0151 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5691744Z   triton_flex_attention_1998 0.0160 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5692375Z   triton_flex_attention_1978 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5692506Z SingleProcess AUTOTUNE benchmarking takes 0.2059 seconds and 0.3341 seconds precompiling for 24 choices
2025-12-04T09:45:17.5692546Z Autotune Choices Stats:
2025-12-04T09:45:17.5693310Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2019", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.5693546Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5693723Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5693999Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5694630Z   triton_flex_attention_backward_2019 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5695246Z   triton_flex_attention_backward_2013 0.0210 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5695875Z   triton_flex_attention_backward_2010 0.0214 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5696505Z   triton_flex_attention_backward_2011 0.0214 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5697142Z   triton_flex_attention_backward_2021 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5697775Z   triton_flex_attention_backward_2020 0.0233 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5698412Z   triton_flex_attention_backward_2018 0.0250 ms 72.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5699034Z   triton_flex_attention_backward_2023 0.0253 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5699657Z   triton_flex_attention_backward_2014 0.0262 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5700280Z   triton_flex_attention_backward_2005 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5700453Z SingleProcess AUTOTUNE benchmarking takes 0.2422 seconds and 0.7502 seconds precompiling for 22 choices
2025-12-04T09:45:17.5700532Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5700574Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5700614Z unimplemented []
2025-12-04T09:45:17.5700677Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5700794Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5701368Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5701420Z graph_break []
2025-12-04T09:45:17.5701494Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5702751Z Autotune Choices Stats:
2025-12-04T09:45:17.5703502Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2030", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.5703633Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5703751Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5703916Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5704528Z   triton_flex_attention_2030 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5705130Z   triton_flex_attention_2028 0.0109 ms 94.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5705737Z   triton_flex_attention_2031 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5706357Z   triton_flex_attention_2026 0.0126 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5706959Z   triton_flex_attention_2029 0.0127 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5707598Z   triton_flex_attention_2027 0.0142 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5708220Z   triton_flex_attention_2046 0.0147 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5708825Z   triton_flex_attention_2038 0.0152 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5709433Z   triton_flex_attention_2044 0.0162 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5710031Z   triton_flex_attention_2024 0.0165 ms 62.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5710162Z SingleProcess AUTOTUNE benchmarking takes 0.2047 seconds and 0.3631 seconds precompiling for 24 choices
2025-12-04T09:45:17.5710205Z Autotune Choices Stats:
2025-12-04T09:45:17.5710997Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2065", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017799999564886093, "best_triton_pos": 0}
2025-12-04T09:45:17.5711216Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5711396Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5711687Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5712322Z   triton_flex_attention_backward_2065 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5712952Z   triton_flex_attention_backward_2059 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5713573Z   triton_flex_attention_backward_2056 0.0213 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5714210Z   triton_flex_attention_backward_2057 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5714847Z   triton_flex_attention_backward_2067 0.0230 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5715485Z   triton_flex_attention_backward_2066 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5716111Z   triton_flex_attention_backward_2064 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5716756Z   triton_flex_attention_backward_2069 0.0252 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5717386Z   triton_flex_attention_backward_2060 0.0260 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5718008Z   triton_flex_attention_backward_2051 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5718138Z SingleProcess AUTOTUNE benchmarking takes 0.2494 seconds and 0.8153 seconds precompiling for 22 choices
2025-12-04T09:45:17.5718214Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5718257Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5718299Z unimplemented []
2025-12-04T09:45:17.5718362Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5718466Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5719045Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5719089Z graph_break []
2025-12-04T09:45:17.5719164Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5719207Z Autotune Choices Stats:
2025-12-04T09:45:17.5719948Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2076", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.5720094Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5720210Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5720385Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5721028Z   triton_flex_attention_2076 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5721636Z   triton_flex_attention_2074 0.0108 ms 94.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5722242Z   triton_flex_attention_2077 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5722839Z   triton_flex_attention_2072 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5723466Z   triton_flex_attention_2075 0.0125 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5724068Z   triton_flex_attention_2073 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5724714Z   triton_flex_attention_2092 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5725319Z   triton_flex_attention_2084 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5725927Z   triton_flex_attention_2090 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5726532Z   triton_flex_attention_2070 0.0167 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5726663Z SingleProcess AUTOTUNE benchmarking takes 0.2086 seconds and 0.3462 seconds precompiling for 24 choices
2025-12-04T09:45:17.5726703Z Autotune Choices Stats:
2025-12-04T09:45:17.5727465Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2111", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017680000513792038, "best_triton_pos": 0}
2025-12-04T09:45:17.5727682Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5727850Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5728131Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5728770Z   triton_flex_attention_backward_2111 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5729412Z   triton_flex_attention_backward_2105 0.0210 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5730037Z   triton_flex_attention_backward_2102 0.0214 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5730685Z   triton_flex_attention_backward_2103 0.0215 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5731312Z   triton_flex_attention_backward_2113 0.0232 ms 76.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5731948Z   triton_flex_attention_backward_2112 0.0234 ms 75.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5732646Z   triton_flex_attention_backward_2110 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5733312Z   triton_flex_attention_backward_2115 0.0253 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5733934Z   triton_flex_attention_backward_2106 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5734558Z   triton_flex_attention_backward_2097 0.0262 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5734688Z SingleProcess AUTOTUNE benchmarking takes 0.2473 seconds and 0.8010 seconds precompiling for 22 choices
2025-12-04T09:45:17.5734761Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5734807Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5734845Z unimplemented []
2025-12-04T09:45:17.5734909Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5735008Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5735572Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5735611Z graph_break []
2025-12-04T09:45:17.5735684Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5735730Z Autotune Choices Stats:
2025-12-04T09:45:17.5736468Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2122", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.5736595Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5736718Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5736892Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5737542Z   triton_flex_attention_2122 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5738147Z   triton_flex_attention_2120 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5738749Z   triton_flex_attention_2123 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5739353Z   triton_flex_attention_2118 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5739954Z   triton_flex_attention_2121 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5740589Z   triton_flex_attention_2119 0.0142 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5741194Z   triton_flex_attention_2138 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5741834Z   triton_flex_attention_2130 0.0151 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5742441Z   triton_flex_attention_2136 0.0163 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5743045Z   triton_flex_attention_2116 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5743175Z SingleProcess AUTOTUNE benchmarking takes 0.2130 seconds and 0.3464 seconds precompiling for 24 choices
2025-12-04T09:45:17.5743217Z Autotune Choices Stats:
2025-12-04T09:45:17.5743981Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2157", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.5744200Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5744378Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5744654Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5745283Z   triton_flex_attention_backward_2157 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5745938Z   triton_flex_attention_backward_2151 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5746560Z   triton_flex_attention_backward_2148 0.0217 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5747178Z   triton_flex_attention_backward_2149 0.0217 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5747801Z   triton_flex_attention_backward_2159 0.0234 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5748422Z   triton_flex_attention_backward_2158 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5749053Z   triton_flex_attention_backward_2156 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5749678Z   triton_flex_attention_backward_2161 0.0256 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5750325Z   triton_flex_attention_backward_2152 0.0261 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5750971Z   triton_flex_attention_backward_2143 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5751099Z SingleProcess AUTOTUNE benchmarking takes 0.2464 seconds and 0.8851 seconds precompiling for 22 choices
2025-12-04T09:45:17.5751173Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5751217Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5751255Z unimplemented []
2025-12-04T09:45:17.5751315Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5751415Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5751988Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5752027Z graph_break []
2025-12-04T09:45:17.5752101Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5752142Z Autotune Choices Stats:
2025-12-04T09:45:17.5752888Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2168", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009840000420808792, "best_triton_pos": 0}
2025-12-04T09:45:17.5753015Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5753129Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5753290Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5753906Z   triton_flex_attention_2168 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5754551Z   triton_flex_attention_2166 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5755147Z   triton_flex_attention_2169 0.0114 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5755746Z   triton_flex_attention_2167 0.0124 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5756347Z   triton_flex_attention_2164 0.0124 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5756946Z   triton_flex_attention_2165 0.0145 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5757562Z   triton_flex_attention_2184 0.0146 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5758164Z   triton_flex_attention_2176 0.0150 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5758795Z   triton_flex_attention_2182 0.0160 ms 61.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5759396Z   triton_flex_attention_2174 0.0167 ms 58.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5759525Z SingleProcess AUTOTUNE benchmarking takes 0.2149 seconds and 0.3567 seconds precompiling for 24 choices
2025-12-04T09:45:17.5759567Z Autotune Choices Stats:
2025-12-04T09:45:17.5760329Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2203", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.5760578Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5760745Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5761024Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5761674Z   triton_flex_attention_backward_2203 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5762297Z   triton_flex_attention_backward_2197 0.0210 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5762954Z   triton_flex_attention_backward_2194 0.0213 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5763573Z   triton_flex_attention_backward_2195 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5764198Z   triton_flex_attention_backward_2205 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5764816Z   triton_flex_attention_backward_2204 0.0233 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5765437Z   triton_flex_attention_backward_2202 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5766075Z   triton_flex_attention_backward_2207 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5766691Z   triton_flex_attention_backward_2198 0.0262 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5767338Z   triton_flex_attention_backward_2189 0.0266 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5767467Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.8512 seconds precompiling for 22 choices
2025-12-04T09:45:17.5767542Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5767586Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5767624Z unimplemented []
2025-12-04T09:45:17.5767687Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5767787Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5768359Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5768399Z graph_break []
2025-12-04T09:45:17.5768472Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5768514Z Autotune Choices Stats:
2025-12-04T09:45:17.5769252Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2214", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.5769377Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5769491Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5769666Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5770280Z   triton_flex_attention_2214 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5770926Z   triton_flex_attention_2212 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5771552Z   triton_flex_attention_2215 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5772153Z   triton_flex_attention_2210 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5772762Z   triton_flex_attention_2213 0.0124 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5773370Z   triton_flex_attention_2211 0.0144 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5773983Z   triton_flex_attention_2230 0.0148 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5774609Z   triton_flex_attention_2222 0.0151 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5775213Z   triton_flex_attention_2228 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5775845Z   triton_flex_attention_2208 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5775975Z SingleProcess AUTOTUNE benchmarking takes 0.2066 seconds and 0.3920 seconds precompiling for 24 choices
2025-12-04T09:45:17.5776016Z Autotune Choices Stats:
2025-12-04T09:45:17.5776764Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2249", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.5776981Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5777145Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5777430Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5778065Z   triton_flex_attention_backward_2249 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5778703Z   triton_flex_attention_backward_2243 0.0210 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5779325Z   triton_flex_attention_backward_2241 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5779974Z   triton_flex_attention_backward_2240 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5780642Z   triton_flex_attention_backward_2250 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5781272Z   triton_flex_attention_backward_2251 0.0231 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5781893Z   triton_flex_attention_backward_2248 0.0251 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5782523Z   triton_flex_attention_backward_2253 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5783164Z   triton_flex_attention_backward_2244 0.0261 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5783786Z   triton_flex_attention_backward_2235 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5783941Z SingleProcess AUTOTUNE benchmarking takes 0.2484 seconds and 0.7948 seconds precompiling for 22 choices
2025-12-04T09:45:17.5784035Z ___________ TestLearnableBiasesCUDA.test_flex_attention_logging_cuda ___________
2025-12-04T09:45:17.5784097Z Traceback (most recent call last):
2025-12-04T09:45:17.5784254Z   File "/var/lib/jenkins/pytorch/test/inductor/test_flex_attention.py", line 7343, in test_flex_attention_logging
2025-12-04T09:45:17.5784296Z     self.assertTrue(
2025-12-04T09:45:17.5784402Z   File "/opt/conda/envs/py_3.12/lib/python3.12/unittest/case.py", line 727, in assertTrue
2025-12-04T09:45:17.5784452Z     raise self.failureException(msg)
2025-12-04T09:45:17.5784581Z AssertionError: False is not true : Log file /tmp/tmpwj1h5tyv/flex_attention_configs.json was not created
2025-12-04T09:45:17.5784585Z 
2025-12-04T09:45:17.5784663Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.5784830Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.5784833Z 
2025-12-04T09:45:17.5784923Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.5785000Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5785045Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5785083Z unimplemented []
2025-12-04T09:45:17.5785145Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5785726Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('select_algorithm_num_precompiles', 46), ('async_compile_cache_miss', 33), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2), ('async_compile_cache_hit', 1)]
2025-12-04T09:45:17.5785825Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5785862Z graph_break []
2025-12-04T09:45:17.5785937Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5786427Z /opt/conda/envs/py_3.12/lib/python3.12/site-packages/torch/_inductor/select_algorithm.py:3433: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly.  To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
2025-12-04T09:45:17.5786478Z   current_size = base.storage().size()
2025-12-04T09:45:17.5786519Z Autotune Choices Stats:
2025-12-04T09:45:17.5787281Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_6", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.5787421Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5787535Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5787710Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5788340Z   triton_flex_attention_6 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5788940Z   triton_flex_attention_4 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5789541Z   triton_flex_attention_7 0.0115 ms 88.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5790144Z   triton_flex_attention_2 0.0122 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5790779Z   triton_flex_attention_5 0.0125 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5791388Z   triton_flex_attention_3 0.0145 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5791993Z   triton_flex_attention_22 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5792627Z   triton_flex_attention_14 0.0153 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5793220Z   triton_flex_attention_20 0.0160 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5793828Z   triton_flex_attention_12 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5793960Z SingleProcess AUTOTUNE benchmarking takes 0.1200 seconds and 0.6070 seconds precompiling for 24 choices
2025-12-04T09:45:17.5794002Z Autotune Choices Stats:
2025-12-04T09:45:17.5794760Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_41", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.5794978Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5795154Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5795430Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5796054Z   triton_flex_attention_backward_41 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5796702Z   triton_flex_attention_backward_35 0.0213 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5797326Z   triton_flex_attention_backward_32 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5797965Z   triton_flex_attention_backward_33 0.0216 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5798589Z   triton_flex_attention_backward_42 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5799212Z   triton_flex_attention_backward_43 0.0233 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5799839Z   triton_flex_attention_backward_40 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5800477Z   triton_flex_attention_backward_45 0.0254 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5801141Z   triton_flex_attention_backward_36 0.0264 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5801758Z   triton_flex_attention_backward_27 0.0267 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5801887Z SingleProcess AUTOTUNE benchmarking takes 0.2614 seconds and 0.7025 seconds precompiling for 22 choices
2025-12-04T09:45:17.5801963Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5802006Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5802046Z unimplemented []
2025-12-04T09:45:17.5802107Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5802208Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5802781Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5802821Z graph_break []
2025-12-04T09:45:17.5802894Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5802937Z Autotune Choices Stats:
2025-12-04T09:45:17.5803690Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_52", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.5803818Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5803935Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5804095Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5804706Z   triton_flex_attention_52 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5805332Z   triton_flex_attention_50 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5805927Z   triton_flex_attention_53 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5806530Z   triton_flex_attention_48 0.0122 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5807129Z   triton_flex_attention_51 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5807733Z   triton_flex_attention_49 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5808354Z   triton_flex_attention_68 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5808953Z   triton_flex_attention_60 0.0153 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5809582Z   triton_flex_attention_66 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5810181Z   triton_flex_attention_46 0.0168 ms 59.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5810311Z SingleProcess AUTOTUNE benchmarking takes 0.1997 seconds and 0.3209 seconds precompiling for 24 choices
2025-12-04T09:45:17.5810354Z Autotune Choices Stats:
2025-12-04T09:45:17.5811143Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_87", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.5811359Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5811526Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5811801Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5812438Z   triton_flex_attention_backward_87 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5813056Z   triton_flex_attention_backward_81 0.0209 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5813711Z   triton_flex_attention_backward_78 0.0213 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5814327Z   triton_flex_attention_backward_79 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5814954Z   triton_flex_attention_backward_89 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5815572Z   triton_flex_attention_backward_88 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5816189Z   triton_flex_attention_backward_86 0.0250 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5816822Z   triton_flex_attention_backward_91 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5817443Z   triton_flex_attention_backward_82 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5818097Z   triton_flex_attention_backward_73 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5818227Z SingleProcess AUTOTUNE benchmarking takes 0.2491 seconds and 0.7936 seconds precompiling for 22 choices
2025-12-04T09:45:17.5818300Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5818344Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5818383Z unimplemented []
2025-12-04T09:45:17.5818446Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5818546Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5819126Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5819164Z graph_break []
2025-12-04T09:45:17.5819238Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5819278Z Autotune Choices Stats:
2025-12-04T09:45:17.5820021Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_98", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.5820148Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5820262Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5820464Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5821077Z   triton_flex_attention_98 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5821689Z   triton_flex_attention_96 0.0116 ms 90.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5822311Z   triton_flex_attention_99 0.0118 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5822908Z   triton_flex_attention_94 0.0126 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5823506Z   triton_flex_attention_97 0.0127 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5824112Z   triton_flex_attention_114 0.0146 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5824709Z   triton_flex_attention_95 0.0147 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5825323Z   triton_flex_attention_106 0.0151 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5825923Z   triton_flex_attention_112 0.0159 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5826541Z   triton_flex_attention_104 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5826670Z SingleProcess AUTOTUNE benchmarking takes 0.2065 seconds and 0.3611 seconds precompiling for 24 choices
2025-12-04T09:45:17.5826711Z Autotune Choices Stats:
2025-12-04T09:45:17.5827479Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_133", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.5827694Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5827863Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5828147Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5828784Z   triton_flex_attention_backward_133 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5829412Z   triton_flex_attention_backward_127 0.0212 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5830023Z   triton_flex_attention_backward_124 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5830716Z   triton_flex_attention_backward_125 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5831332Z   triton_flex_attention_backward_134 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5831963Z   triton_flex_attention_backward_135 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5832596Z   triton_flex_attention_backward_137 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5833225Z   triton_flex_attention_backward_132 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5833859Z   triton_flex_attention_backward_128 0.0260 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5834480Z   triton_flex_attention_backward_119 0.0268 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5834640Z SingleProcess AUTOTUNE benchmarking takes 0.2382 seconds and 0.6573 seconds precompiling for 22 choices
2025-12-04T09:45:17.5834715Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5834767Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5834806Z unimplemented []
2025-12-04T09:45:17.5834866Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5834966Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5835538Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5835577Z graph_break []
2025-12-04T09:45:17.5835650Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5835690Z Autotune Choices Stats:
2025-12-04T09:45:17.5836425Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_144", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010639999993145466, "best_triton_pos": 0}
2025-12-04T09:45:17.5836553Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5836668Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5836830Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5837450Z   triton_flex_attention_144 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5838052Z   triton_flex_attention_142 0.0113 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5838659Z   triton_flex_attention_145 0.0116 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5839279Z   triton_flex_attention_140 0.0126 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5839877Z   triton_flex_attention_143 0.0126 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5840508Z   triton_flex_attention_141 0.0141 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5841118Z   triton_flex_attention_160 0.0150 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5841733Z   triton_flex_attention_152 0.0152 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5842333Z   triton_flex_attention_158 0.0162 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5842932Z   triton_flex_attention_150 0.0168 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5843085Z SingleProcess AUTOTUNE benchmarking takes 0.2220 seconds and 0.3273 seconds precompiling for 24 choices
2025-12-04T09:45:17.5843127Z Autotune Choices Stats:
2025-12-04T09:45:17.5843890Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_179", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.5844108Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5844275Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5844550Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5845180Z   triton_flex_attention_backward_179 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5845804Z   triton_flex_attention_backward_173 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5846434Z   triton_flex_attention_backward_170 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5847051Z   triton_flex_attention_backward_171 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5847693Z   triton_flex_attention_backward_181 0.0232 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5848319Z   triton_flex_attention_backward_180 0.0233 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5848938Z   triton_flex_attention_backward_178 0.0251 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5849563Z   triton_flex_attention_backward_183 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5850193Z   triton_flex_attention_backward_174 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5853189Z   triton_flex_attention_backward_165 0.0268 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5853357Z SingleProcess AUTOTUNE benchmarking takes 0.2538 seconds and 0.6741 seconds precompiling for 22 choices
2025-12-04T09:45:17.5853439Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5853499Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5853542Z unimplemented []
2025-12-04T09:45:17.5853607Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5853710Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5854304Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5854345Z graph_break []
2025-12-04T09:45:17.5854424Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5854468Z Autotune Choices Stats:
2025-12-04T09:45:17.5855219Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_190", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010599000379443169, "best_triton_pos": 0}
2025-12-04T09:45:17.5855350Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5855469Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5855635Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5856250Z   triton_flex_attention_190 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5856862Z   triton_flex_attention_188 0.0111 ms 95.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5857464Z   triton_flex_attention_191 0.0114 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5858072Z   triton_flex_attention_186 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5858685Z   triton_flex_attention_189 0.0124 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5859283Z   triton_flex_attention_187 0.0145 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5859886Z   triton_flex_attention_206 0.0149 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5860512Z   triton_flex_attention_198 0.0154 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5861127Z   triton_flex_attention_204 0.0162 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5861722Z   triton_flex_attention_184 0.0169 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5861864Z SingleProcess AUTOTUNE benchmarking takes 0.2042 seconds and 0.3284 seconds precompiling for 24 choices
2025-12-04T09:45:17.5861906Z Autotune Choices Stats:
2025-12-04T09:45:17.5862687Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_225", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.5862907Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5863077Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5863358Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5863990Z   triton_flex_attention_backward_225 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5864636Z   triton_flex_attention_backward_219 0.0211 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5865285Z   triton_flex_attention_backward_217 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5865909Z   triton_flex_attention_backward_216 0.0215 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5866543Z   triton_flex_attention_backward_227 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5867193Z   triton_flex_attention_backward_226 0.0232 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5867810Z   triton_flex_attention_backward_224 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5868456Z   triton_flex_attention_backward_229 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5869105Z   triton_flex_attention_backward_220 0.0261 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5869733Z   triton_flex_attention_backward_211 0.0267 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5869866Z SingleProcess AUTOTUNE benchmarking takes 0.2384 seconds and 0.6973 seconds precompiling for 22 choices
2025-12-04T09:45:17.5869943Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5869988Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5870027Z unimplemented []
2025-12-04T09:45:17.5870090Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5870190Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5870813Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5870865Z graph_break []
2025-12-04T09:45:17.5870937Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5870978Z Autotune Choices Stats:
2025-12-04T09:45:17.5871728Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_236", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009399999864399433, "best_triton_pos": 0}
2025-12-04T09:45:17.5871859Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5871975Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5872135Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5872757Z   triton_flex_attention_236 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5873363Z   triton_flex_attention_234 0.0108 ms 87.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5873976Z   triton_flex_attention_237 0.0114 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5874574Z   triton_flex_attention_232 0.0122 ms 77.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5875178Z   triton_flex_attention_235 0.0124 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5875790Z   triton_flex_attention_233 0.0144 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5876398Z   triton_flex_attention_252 0.0145 ms 64.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5877007Z   triton_flex_attention_244 0.0151 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5877610Z   triton_flex_attention_250 0.0160 ms 58.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5878216Z   triton_flex_attention_242 0.0167 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5878347Z SingleProcess AUTOTUNE benchmarking takes 0.2039 seconds and 0.3319 seconds precompiling for 24 choices
2025-12-04T09:45:17.5878389Z Autotune Choices Stats:
2025-12-04T09:45:17.5879138Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_271", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.5879376Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5879552Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5879828Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5880487Z   triton_flex_attention_backward_271 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5881108Z   triton_flex_attention_backward_265 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5881732Z   triton_flex_attention_backward_262 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5882369Z   triton_flex_attention_backward_263 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5882991Z   triton_flex_attention_backward_273 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5883629Z   triton_flex_attention_backward_272 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5884277Z   triton_flex_attention_backward_270 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5884902Z   triton_flex_attention_backward_275 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5885540Z   triton_flex_attention_backward_257 0.0262 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5886165Z   triton_flex_attention_backward_266 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5886293Z SingleProcess AUTOTUNE benchmarking takes 0.2642 seconds and 0.6684 seconds precompiling for 22 choices
2025-12-04T09:45:17.5886368Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5886410Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5886450Z unimplemented []
2025-12-04T09:45:17.5886521Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5886622Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5887196Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5887243Z graph_break []
2025-12-04T09:45:17.5887317Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5887368Z Autotune Choices Stats:
2025-12-04T09:45:17.5888126Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_282", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010119999758899212, "best_triton_pos": 0}
2025-12-04T09:45:17.5888252Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5888368Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5888531Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5889140Z   triton_flex_attention_282 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5889741Z   triton_flex_attention_280 0.0110 ms 92.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5890344Z   triton_flex_attention_283 0.0111 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5890986Z   triton_flex_attention_278 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5891584Z   triton_flex_attention_281 0.0123 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5892207Z   triton_flex_attention_279 0.0143 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5892827Z   triton_flex_attention_298 0.0147 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5893427Z   triton_flex_attention_290 0.0153 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5894019Z   triton_flex_attention_296 0.0162 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5894621Z   triton_flex_attention_276 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5894751Z SingleProcess AUTOTUNE benchmarking takes 0.2005 seconds and 0.3169 seconds precompiling for 24 choices
2025-12-04T09:45:17.5894791Z Autotune Choices Stats:
2025-12-04T09:45:17.5895548Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_317", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.5895764Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5895940Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5896226Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5896866Z   triton_flex_attention_backward_317 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5897481Z   triton_flex_attention_backward_311 0.0211 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5898102Z   triton_flex_attention_backward_308 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5898717Z   triton_flex_attention_backward_309 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5899353Z   triton_flex_attention_backward_319 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5899976Z   triton_flex_attention_backward_318 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5900635Z   triton_flex_attention_backward_316 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5901287Z   triton_flex_attention_backward_321 0.0254 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5901909Z   triton_flex_attention_backward_312 0.0263 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5902529Z   triton_flex_attention_backward_303 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5902659Z SingleProcess AUTOTUNE benchmarking takes 0.2394 seconds and 0.8193 seconds precompiling for 22 choices
2025-12-04T09:45:17.5902732Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5902778Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5902818Z unimplemented []
2025-12-04T09:45:17.5902880Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5902979Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5903570Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5903610Z graph_break []
2025-12-04T09:45:17.5903682Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5903724Z Autotune Choices Stats:
2025-12-04T09:45:17.5904460Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_328", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009359999559819698, "best_triton_pos": 0}
2025-12-04T09:45:17.5904612Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5904724Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5904895Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5905507Z   triton_flex_attention_328 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5906112Z   triton_flex_attention_326 0.0108 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5906713Z   triton_flex_attention_329 0.0110 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5907315Z   triton_flex_attention_324 0.0120 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5907925Z   triton_flex_attention_327 0.0126 ms 74.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5908515Z   triton_flex_attention_325 0.0140 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5909145Z   triton_flex_attention_344 0.0145 ms 64.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5909746Z   triton_flex_attention_336 0.0151 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5910349Z   triton_flex_attention_342 0.0160 ms 58.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5910994Z   triton_flex_attention_334 0.0166 ms 56.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5911124Z SingleProcess AUTOTUNE benchmarking takes 0.2054 seconds and 0.4299 seconds precompiling for 24 choices
2025-12-04T09:45:17.5911166Z Autotune Choices Stats:
2025-12-04T09:45:17.5911929Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_363", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.5912148Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5912317Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5912597Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5913236Z   triton_flex_attention_backward_363 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5913882Z   triton_flex_attention_backward_357 0.0211 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5914504Z   triton_flex_attention_backward_355 0.0214 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5915123Z   triton_flex_attention_backward_354 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5915766Z   triton_flex_attention_backward_365 0.0230 ms 78.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5916409Z   triton_flex_attention_backward_364 0.0232 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5917030Z   triton_flex_attention_backward_362 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5917686Z   triton_flex_attention_backward_367 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5918308Z   triton_flex_attention_backward_358 0.0262 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5918950Z   triton_flex_attention_backward_349 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5919077Z SingleProcess AUTOTUNE benchmarking takes 0.2330 seconds and 0.6710 seconds precompiling for 22 choices
2025-12-04T09:45:17.5919153Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5919196Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5919235Z unimplemented []
2025-12-04T09:45:17.5919297Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5919397Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5919970Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5920010Z graph_break []
2025-12-04T09:45:17.5920085Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5920127Z Autotune Choices Stats:
2025-12-04T09:45:17.5920915Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_374", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.5921043Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5921173Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5921346Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5921981Z   triton_flex_attention_374 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5922585Z   triton_flex_attention_372 0.0106 ms 96.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5923188Z   triton_flex_attention_375 0.0113 ms 91.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5923808Z   triton_flex_attention_370 0.0123 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5924410Z   triton_flex_attention_373 0.0125 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5925019Z   triton_flex_attention_371 0.0144 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5925622Z   triton_flex_attention_390 0.0149 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5926248Z   triton_flex_attention_382 0.0153 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5926849Z   triton_flex_attention_388 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5927447Z   triton_flex_attention_368 0.0167 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5927579Z SingleProcess AUTOTUNE benchmarking takes 0.2071 seconds and 0.3442 seconds precompiling for 24 choices
2025-12-04T09:45:17.5927620Z Autotune Choices Stats:
2025-12-04T09:45:17.5928369Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_409", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.5928585Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5928764Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5929038Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5929665Z   triton_flex_attention_backward_409 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5930321Z   triton_flex_attention_backward_403 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5930970Z   triton_flex_attention_backward_400 0.0214 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5931592Z   triton_flex_attention_backward_401 0.0214 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5932210Z   triton_flex_attention_backward_411 0.0229 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5932832Z   triton_flex_attention_backward_410 0.0232 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5933459Z   triton_flex_attention_backward_408 0.0251 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5934082Z   triton_flex_attention_backward_413 0.0252 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5934741Z   triton_flex_attention_backward_404 0.0263 ms 69.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5935362Z   triton_flex_attention_backward_395 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5935493Z SingleProcess AUTOTUNE benchmarking takes 0.2510 seconds and 0.7196 seconds precompiling for 22 choices
2025-12-04T09:45:17.5935566Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5935609Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5935649Z unimplemented []
2025-12-04T09:45:17.5935711Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5935810Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5936379Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.5936419Z graph_break []
2025-12-04T09:45:17.5936494Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5936535Z Autotune Choices Stats:
2025-12-04T09:45:17.5937277Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_420", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01056000031530857, "best_triton_pos": 0}
2025-12-04T09:45:17.5937406Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5937519Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5937681Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5938282Z   triton_flex_attention_420 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5938916Z   triton_flex_attention_418 0.0109 ms 96.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5939515Z   triton_flex_attention_421 0.0113 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5940136Z   triton_flex_attention_416 0.0122 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5940765Z   triton_flex_attention_419 0.0123 ms 86.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5941366Z   triton_flex_attention_417 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5941981Z   triton_flex_attention_436 0.0146 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5942582Z   triton_flex_attention_428 0.0150 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5943220Z   triton_flex_attention_434 0.0160 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5943816Z   triton_flex_attention_426 0.0166 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5943946Z SingleProcess AUTOTUNE benchmarking takes 0.2081 seconds and 0.3328 seconds precompiling for 24 choices
2025-12-04T09:45:17.5943987Z Autotune Choices Stats:
2025-12-04T09:45:17.5944741Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_455", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.5944959Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5945125Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5945404Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5946062Z   triton_flex_attention_backward_455 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5946684Z   triton_flex_attention_backward_449 0.0208 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5947336Z   triton_flex_attention_backward_446 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5947959Z   triton_flex_attention_backward_447 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5948586Z   triton_flex_attention_backward_457 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5949214Z   triton_flex_attention_backward_456 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5949823Z   triton_flex_attention_backward_454 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5950483Z   triton_flex_attention_backward_459 0.0254 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5951104Z   triton_flex_attention_backward_450 0.0262 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5951756Z   triton_flex_attention_backward_441 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5951884Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.6920 seconds precompiling for 22 choices
2025-12-04T09:45:17.5951962Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5952005Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5952045Z unimplemented []
2025-12-04T09:45:17.5952106Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5952206Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5952780Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5952820Z graph_break []
2025-12-04T09:45:17.5952893Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5952936Z Autotune Choices Stats:
2025-12-04T09:45:17.5953679Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_466", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01011900044977665, "best_triton_pos": 0}
2025-12-04T09:45:17.5953805Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5953920Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5954082Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5954708Z   triton_flex_attention_466 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5955307Z   triton_flex_attention_464 0.0112 ms 90.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5955935Z   triton_flex_attention_467 0.0113 ms 89.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5956535Z   triton_flex_attention_462 0.0122 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5957134Z   triton_flex_attention_465 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5957741Z   triton_flex_attention_463 0.0144 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5958349Z   triton_flex_attention_482 0.0148 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5958962Z   triton_flex_attention_474 0.0155 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5959553Z   triton_flex_attention_480 0.0163 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5960178Z   triton_flex_attention_460 0.0167 ms 60.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5960306Z SingleProcess AUTOTUNE benchmarking takes 0.2064 seconds and 0.3251 seconds precompiling for 24 choices
2025-12-04T09:45:17.5960348Z Autotune Choices Stats:
2025-12-04T09:45:17.5961156Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_501", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.5961375Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5961542Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5961820Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5962440Z   triton_flex_attention_backward_501 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5963083Z   triton_flex_attention_backward_495 0.0212 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5963702Z   triton_flex_attention_backward_492 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5964364Z   triton_flex_attention_backward_493 0.0218 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5964988Z   triton_flex_attention_backward_503 0.0234 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5965632Z   triton_flex_attention_backward_502 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5966250Z   triton_flex_attention_backward_500 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5966872Z   triton_flex_attention_backward_505 0.0256 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5967508Z   triton_flex_attention_backward_487 0.0266 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5968125Z   triton_flex_attention_backward_496 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5968274Z SingleProcess AUTOTUNE benchmarking takes 0.2526 seconds and 0.6711 seconds precompiling for 22 choices
2025-12-04T09:45:17.5968347Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5968390Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5968438Z unimplemented []
2025-12-04T09:45:17.5968501Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5968600Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5969164Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 67), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 21), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 5), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5969203Z graph_break []
2025-12-04T09:45:17.5969277Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5969317Z Autotune Choices Stats:
2025-12-04T09:45:17.5970047Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_512", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010239999741315842, "best_triton_pos": 0}
2025-12-04T09:45:17.5970174Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5970287Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5970485Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5971106Z   triton_flex_attention_512 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5971705Z   triton_flex_attention_510 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5972320Z   triton_flex_attention_513 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5972938Z   triton_flex_attention_508 0.0122 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5973537Z   triton_flex_attention_511 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5974136Z   triton_flex_attention_509 0.0143 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5974739Z   triton_flex_attention_528 0.0146 ms 70.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5975341Z   triton_flex_attention_520 0.0151 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5975970Z   triton_flex_attention_526 0.0162 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5976570Z   triton_flex_attention_506 0.0168 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5976720Z SingleProcess AUTOTUNE benchmarking takes 0.2122 seconds and 0.4604 seconds precompiling for 24 choices
2025-12-04T09:45:17.5976761Z Autotune Choices Stats:
2025-12-04T09:45:17.5977527Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_547", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.5977745Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5977911Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5978186Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5978819Z   triton_flex_attention_backward_547 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5979445Z   triton_flex_attention_backward_541 0.0210 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5980102Z   triton_flex_attention_backward_538 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5980759Z   triton_flex_attention_backward_539 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5981424Z   triton_flex_attention_backward_548 0.0232 ms 76.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5982047Z   triton_flex_attention_backward_549 0.0233 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5982664Z   triton_flex_attention_backward_546 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5983288Z   triton_flex_attention_backward_551 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5983908Z   triton_flex_attention_backward_542 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5984548Z   triton_flex_attention_backward_533 0.0267 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5984678Z SingleProcess AUTOTUNE benchmarking takes 0.2462 seconds and 0.8028 seconds precompiling for 22 choices
2025-12-04T09:45:17.5984763Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.5984814Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.5984853Z unimplemented []
2025-12-04T09:45:17.5984913Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.5985013Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.5985597Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.5985637Z graph_break []
2025-12-04T09:45:17.5985710Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.5985751Z Autotune Choices Stats:
2025-12-04T09:45:17.5986480Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_558", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.5986605Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5986721Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5986884Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5987505Z   triton_flex_attention_558 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5988113Z   triton_flex_attention_556 0.0109 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5988712Z   triton_flex_attention_559 0.0112 ms 88.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5989318Z   triton_flex_attention_554 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5989938Z   triton_flex_attention_557 0.0125 ms 79.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5990573Z   triton_flex_attention_555 0.0144 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.5991169Z   triton_flex_attention_574 0.0146 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5991764Z   triton_flex_attention_566 0.0152 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5992391Z   triton_flex_attention_572 0.0160 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5992989Z   triton_flex_attention_564 0.0167 ms 59.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5993115Z SingleProcess AUTOTUNE benchmarking takes 0.2052 seconds and 0.4427 seconds precompiling for 24 choices
2025-12-04T09:45:17.5993173Z Autotune Choices Stats:
2025-12-04T09:45:17.5994044Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_593", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.5994280Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.5994447Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.5994724Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.5995352Z   triton_flex_attention_backward_593 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5995973Z   triton_flex_attention_backward_587 0.0209 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5996595Z   triton_flex_attention_backward_585 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5997226Z   triton_flex_attention_backward_584 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5997842Z   triton_flex_attention_backward_595 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5998491Z   triton_flex_attention_backward_594 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5999114Z   triton_flex_attention_backward_592 0.0249 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.5999739Z   triton_flex_attention_backward_597 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6000373Z   triton_flex_attention_backward_588 0.0260 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6001029Z   triton_flex_attention_backward_579 0.0262 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6001160Z SingleProcess AUTOTUNE benchmarking takes 0.2468 seconds and 0.7679 seconds precompiling for 22 choices
2025-12-04T09:45:17.6001234Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6001277Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6001316Z unimplemented []
2025-12-04T09:45:17.6001380Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6001479Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6002066Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.6002116Z graph_break []
2025-12-04T09:45:17.6002190Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6002230Z Autotune Choices Stats:
2025-12-04T09:45:17.6002980Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_604", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009998999536037445, "best_triton_pos": 0}
2025-12-04T09:45:17.6003109Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6003223Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6003386Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6003997Z   triton_flex_attention_604 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6004601Z   triton_flex_attention_602 0.0105 ms 95.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6005214Z   triton_flex_attention_605 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6005816Z   triton_flex_attention_600 0.0122 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6006422Z   triton_flex_attention_603 0.0124 ms 80.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6007038Z   triton_flex_attention_601 0.0143 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6007645Z   triton_flex_attention_620 0.0147 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6008245Z   triton_flex_attention_612 0.0152 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6008867Z   triton_flex_attention_618 0.0160 ms 62.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6009498Z   triton_flex_attention_598 0.0166 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6009628Z SingleProcess AUTOTUNE benchmarking takes 0.2062 seconds and 0.3317 seconds precompiling for 24 choices
2025-12-04T09:45:17.6009668Z Autotune Choices Stats:
2025-12-04T09:45:17.6010455Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_639", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.6010704Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6010883Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6011160Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6011792Z   triton_flex_attention_backward_639 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6012418Z   triton_flex_attention_backward_633 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6013033Z   triton_flex_attention_backward_630 0.0216 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6013668Z   triton_flex_attention_backward_631 0.0216 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6014298Z   triton_flex_attention_backward_641 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6014931Z   triton_flex_attention_backward_640 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6015567Z   triton_flex_attention_backward_638 0.0250 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6016192Z   triton_flex_attention_backward_643 0.0254 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6016816Z   triton_flex_attention_backward_634 0.0262 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6017438Z   triton_flex_attention_backward_625 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6017572Z SingleProcess AUTOTUNE benchmarking takes 0.2408 seconds and 0.7983 seconds precompiling for 22 choices
2025-12-04T09:45:17.6017648Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6017690Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6017729Z unimplemented []
2025-12-04T09:45:17.6017789Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6017900Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6018471Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6018520Z graph_break []
2025-12-04T09:45:17.6018594Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6018636Z Autotune Choices Stats:
2025-12-04T09:45:17.6019384Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_650", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009440000168979168, "best_triton_pos": 0}
2025-12-04T09:45:17.6019512Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6019627Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6019788Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6020397Z   triton_flex_attention_650 0.0094 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6021034Z   triton_flex_attention_648 0.0107 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6021634Z   triton_flex_attention_651 0.0113 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6022247Z   triton_flex_attention_649 0.0123 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6022846Z   triton_flex_attention_646 0.0125 ms 75.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6023456Z   triton_flex_attention_647 0.0141 ms 66.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6024081Z   triton_flex_attention_666 0.0148 ms 64.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6024687Z   triton_flex_attention_658 0.0152 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6025291Z   triton_flex_attention_664 0.0162 ms 58.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6025901Z   triton_flex_attention_644 0.0166 ms 56.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6026028Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3296 seconds precompiling for 24 choices
2025-12-04T09:45:17.6026069Z Autotune Choices Stats:
2025-12-04T09:45:17.6026832Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_685", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017959000542759895, "best_triton_pos": 0}
2025-12-04T09:45:17.6027049Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6027231Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6027519Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6028158Z   triton_flex_attention_backward_685 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6028772Z   triton_flex_attention_backward_679 0.0209 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6029391Z   triton_flex_attention_backward_676 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6030016Z   triton_flex_attention_backward_677 0.0217 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6030679Z   triton_flex_attention_backward_687 0.0232 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6031304Z   triton_flex_attention_backward_686 0.0235 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6031935Z   triton_flex_attention_backward_684 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6032578Z   triton_flex_attention_backward_689 0.0255 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6033203Z   triton_flex_attention_backward_680 0.0263 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6033822Z   triton_flex_attention_backward_671 0.0268 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6033950Z SingleProcess AUTOTUNE benchmarking takes 0.2519 seconds and 0.6959 seconds precompiling for 22 choices
2025-12-04T09:45:17.6034025Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6034067Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6034105Z unimplemented []
2025-12-04T09:45:17.6034167Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6034265Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6034850Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6034889Z graph_break []
2025-12-04T09:45:17.6034963Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6035003Z Autotune Choices Stats:
2025-12-04T09:45:17.6035743Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_696", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.6035890Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6036002Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6036173Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6036783Z   triton_flex_attention_696 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6037386Z   triton_flex_attention_694 0.0110 ms 89.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6037986Z   triton_flex_attention_697 0.0112 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6038607Z   triton_flex_attention_692 0.0120 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6039214Z   triton_flex_attention_695 0.0126 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6039814Z   triton_flex_attention_693 0.0142 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6040475Z   triton_flex_attention_712 0.0146 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6041094Z   triton_flex_attention_704 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6041694Z   triton_flex_attention_710 0.0162 ms 61.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6042292Z   triton_flex_attention_702 0.0167 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6042427Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.3438 seconds precompiling for 24 choices
2025-12-04T09:45:17.6042466Z Autotune Choices Stats:
2025-12-04T09:45:17.6043228Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_731", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.6043465Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6043630Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6043909Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6044547Z   triton_flex_attention_backward_731 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6045193Z   triton_flex_attention_backward_725 0.0211 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6045813Z   triton_flex_attention_backward_722 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6046432Z   triton_flex_attention_backward_723 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6047079Z   triton_flex_attention_backward_733 0.0232 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6047791Z   triton_flex_attention_backward_732 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6048409Z   triton_flex_attention_backward_730 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6049055Z   triton_flex_attention_backward_735 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6049688Z   triton_flex_attention_backward_726 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6050306Z   triton_flex_attention_backward_717 0.0264 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6050470Z SingleProcess AUTOTUNE benchmarking takes 0.2420 seconds and 0.8787 seconds precompiling for 22 choices
2025-12-04T09:45:17.6050545Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6050588Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6050629Z unimplemented []
2025-12-04T09:45:17.6050690Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6050791Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6051359Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.6051397Z graph_break []
2025-12-04T09:45:17.6051470Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6051510Z Autotune Choices Stats:
2025-12-04T09:45:17.6052267Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_742", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.6052393Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6052520Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6052691Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6053336Z   triton_flex_attention_742 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6053935Z   triton_flex_attention_740 0.0109 ms 91.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6054537Z   triton_flex_attention_743 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6055131Z   triton_flex_attention_738 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6055732Z   triton_flex_attention_741 0.0126 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6056345Z   triton_flex_attention_739 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6056949Z   triton_flex_attention_758 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6057579Z   triton_flex_attention_750 0.0152 ms 66.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6058181Z   triton_flex_attention_756 0.0162 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6058782Z   triton_flex_attention_748 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6058910Z SingleProcess AUTOTUNE benchmarking takes 0.2048 seconds and 0.5232 seconds precompiling for 24 choices
2025-12-04T09:45:17.6058953Z Autotune Choices Stats:
2025-12-04T09:45:17.6059710Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_777", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.6059927Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6060095Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6060379Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6061053Z   triton_flex_attention_backward_777 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6061701Z   triton_flex_attention_backward_771 0.0209 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6062328Z   triton_flex_attention_backward_768 0.0214 ms 83.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6062949Z   triton_flex_attention_backward_769 0.0216 ms 82.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6063592Z   triton_flex_attention_backward_779 0.0230 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6064238Z   triton_flex_attention_backward_778 0.0231 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6064866Z   triton_flex_attention_backward_776 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6065495Z   triton_flex_attention_backward_781 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6066143Z   triton_flex_attention_backward_772 0.0260 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6066763Z   triton_flex_attention_backward_763 0.0263 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6066895Z SingleProcess AUTOTUNE benchmarking takes 0.2554 seconds and 0.8189 seconds precompiling for 22 choices
2025-12-04T09:45:17.6066971Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6067013Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6067053Z unimplemented []
2025-12-04T09:45:17.6067115Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6067216Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6067790Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6067828Z graph_break []
2025-12-04T09:45:17.6067902Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6067942Z Autotune Choices Stats:
2025-12-04T09:45:17.6068681Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_788", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009959000162780285, "best_triton_pos": 0}
2025-12-04T09:45:17.6068820Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6068933Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6069098Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6069708Z   triton_flex_attention_788 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6070334Z   triton_flex_attention_786 0.0108 ms 92.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6070969Z   triton_flex_attention_789 0.0114 ms 87.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6071567Z   triton_flex_attention_784 0.0123 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6072183Z   triton_flex_attention_787 0.0126 ms 79.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6072790Z   triton_flex_attention_785 0.0143 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6073406Z   triton_flex_attention_804 0.0145 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6074007Z   triton_flex_attention_796 0.0151 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6074650Z   triton_flex_attention_802 0.0162 ms 61.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6075249Z   triton_flex_attention_782 0.0168 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6075381Z SingleProcess AUTOTUNE benchmarking takes 0.2156 seconds and 0.4496 seconds precompiling for 24 choices
2025-12-04T09:45:17.6075422Z Autotune Choices Stats:
2025-12-04T09:45:17.6076169Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_823", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017719000577926636, "best_triton_pos": 0}
2025-12-04T09:45:17.6076391Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6076556Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6076830Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6077472Z   triton_flex_attention_backward_823 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6078093Z   triton_flex_attention_backward_817 0.0208 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6078740Z   triton_flex_attention_backward_814 0.0215 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6079362Z   triton_flex_attention_backward_815 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6080008Z   triton_flex_attention_backward_825 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6080662Z   triton_flex_attention_backward_824 0.0233 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6081290Z   triton_flex_attention_backward_822 0.0248 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6081933Z   triton_flex_attention_backward_827 0.0253 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6082554Z   triton_flex_attention_backward_818 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6083208Z   triton_flex_attention_backward_809 0.0264 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6083336Z SingleProcess AUTOTUNE benchmarking takes 0.2475 seconds and 0.7974 seconds precompiling for 22 choices
2025-12-04T09:45:17.6083410Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6083453Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6083492Z unimplemented []
2025-12-04T09:45:17.6083553Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6083654Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6084223Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6084263Z graph_break []
2025-12-04T09:45:17.6084337Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6084379Z Autotune Choices Stats:
2025-12-04T09:45:17.6085120Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_834", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.6085250Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6085366Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6085528Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6086145Z   triton_flex_attention_834 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6086743Z   triton_flex_attention_832 0.0110 ms 94.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6087371Z   triton_flex_attention_835 0.0113 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6087972Z   triton_flex_attention_830 0.0121 ms 85.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6088575Z   triton_flex_attention_833 0.0125 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6089170Z   triton_flex_attention_831 0.0143 ms 72.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6089772Z   triton_flex_attention_850 0.0149 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6090377Z   triton_flex_attention_842 0.0154 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6091007Z   triton_flex_attention_848 0.0164 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6091657Z   triton_flex_attention_828 0.0169 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6091787Z SingleProcess AUTOTUNE benchmarking takes 0.2068 seconds and 0.4652 seconds precompiling for 24 choices
2025-12-04T09:45:17.6091829Z Autotune Choices Stats:
2025-12-04T09:45:17.6092584Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_869", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018279999494552612, "best_triton_pos": 0}
2025-12-04T09:45:17.6092803Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6092969Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6093247Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6093876Z   triton_flex_attention_backward_869 0.0183 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6094515Z   triton_flex_attention_backward_863 0.0211 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6095125Z   triton_flex_attention_backward_860 0.0217 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6095774Z   triton_flex_attention_backward_861 0.0217 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6096401Z   triton_flex_attention_backward_870 0.0235 ms 77.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6097027Z   triton_flex_attention_backward_871 0.0236 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6097664Z   triton_flex_attention_backward_868 0.0253 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6098288Z   triton_flex_attention_backward_873 0.0257 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6098919Z   triton_flex_attention_backward_855 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6099541Z   triton_flex_attention_backward_864 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6099691Z SingleProcess AUTOTUNE benchmarking takes 0.2633 seconds and 0.6922 seconds precompiling for 22 choices
2025-12-04T09:45:17.6099768Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6099810Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6099851Z unimplemented []
2025-12-04T09:45:17.6099922Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6100024Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6100631Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6100671Z graph_break []
2025-12-04T09:45:17.6100745Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6100785Z Autotune Choices Stats:
2025-12-04T09:45:17.6101527Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_880", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.6101656Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6101772Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6101938Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6102564Z   triton_flex_attention_880 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6103168Z   triton_flex_attention_881 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6103785Z   triton_flex_attention_878 0.0116 ms 87.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6104415Z   triton_flex_attention_876 0.0122 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6105017Z   triton_flex_attention_879 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6105639Z   triton_flex_attention_877 0.0142 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6106249Z   triton_flex_attention_896 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6106854Z   triton_flex_attention_888 0.0152 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6107477Z   triton_flex_attention_894 0.0160 ms 63.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6108078Z   triton_flex_attention_874 0.0168 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6108226Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.3298 seconds precompiling for 24 choices
2025-12-04T09:45:17.6108266Z Autotune Choices Stats:
2025-12-04T09:45:17.6109035Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_915", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.6109252Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6109421Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6109702Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6110327Z   triton_flex_attention_backward_915 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6110973Z   triton_flex_attention_backward_909 0.0208 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6111599Z   triton_flex_attention_backward_907 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6112221Z   triton_flex_attention_backward_906 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6112876Z   triton_flex_attention_backward_917 0.0230 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6113499Z   triton_flex_attention_backward_916 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6114121Z   triton_flex_attention_backward_914 0.0251 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6114746Z   triton_flex_attention_backward_919 0.0254 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6115368Z   triton_flex_attention_backward_910 0.0262 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6116003Z   triton_flex_attention_backward_901 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6116131Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.6646 seconds precompiling for 22 choices
2025-12-04T09:45:17.6116214Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6116257Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6116307Z unimplemented []
2025-12-04T09:45:17.6116369Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6116469Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6117051Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.6117089Z graph_break []
2025-12-04T09:45:17.6117163Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6117204Z Autotune Choices Stats:
2025-12-04T09:45:17.6117945Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_926", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010359999723732471, "best_triton_pos": 0}
2025-12-04T09:45:17.6118074Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6118189Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6118353Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6118971Z   triton_flex_attention_926 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6119588Z   triton_flex_attention_924 0.0105 ms 98.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6120185Z   triton_flex_attention_927 0.0115 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6120819Z   triton_flex_attention_925 0.0125 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6121447Z   triton_flex_attention_922 0.0127 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6122037Z   triton_flex_attention_923 0.0143 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6122644Z   triton_flex_attention_942 0.0148 ms 70.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6123251Z   triton_flex_attention_934 0.0153 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6123851Z   triton_flex_attention_940 0.0162 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6124468Z   triton_flex_attention_920 0.0167 ms 62.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6124597Z SingleProcess AUTOTUNE benchmarking takes 0.2165 seconds and 0.3269 seconds precompiling for 24 choices
2025-12-04T09:45:17.6124651Z Autotune Choices Stats:
2025-12-04T09:45:17.6125415Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_961", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.6125641Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6125807Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6126086Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6126706Z   triton_flex_attention_backward_961 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6127331Z   triton_flex_attention_backward_955 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6127950Z   triton_flex_attention_backward_952 0.0214 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6128582Z   triton_flex_attention_backward_953 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6129205Z   triton_flex_attention_backward_963 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6129857Z   triton_flex_attention_backward_962 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6130504Z   triton_flex_attention_backward_960 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6131131Z   triton_flex_attention_backward_965 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6131761Z   triton_flex_attention_backward_956 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6132405Z   triton_flex_attention_backward_947 0.0266 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6132533Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.6685 seconds precompiling for 22 choices
2025-12-04T09:45:17.6132606Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6132648Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6132689Z unimplemented []
2025-12-04T09:45:17.6132750Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6132851Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6133426Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 72), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 26), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 10), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6133488Z graph_break []
2025-12-04T09:45:17.6133561Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6133600Z Autotune Choices Stats:
2025-12-04T09:45:17.6134342Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_972", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009560000151395798, "best_triton_pos": 0}
2025-12-04T09:45:17.6134471Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6134585Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6134745Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6135355Z   triton_flex_attention_972 0.0096 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6135963Z   triton_flex_attention_970 0.0110 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6136576Z   triton_flex_attention_973 0.0114 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6137174Z   triton_flex_attention_971 0.0124 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6137780Z   triton_flex_attention_968 0.0126 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6138405Z   triton_flex_attention_969 0.0144 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6139009Z   triton_flex_attention_988 0.0145 ms 66.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6139607Z   triton_flex_attention_980 0.0151 ms 63.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6140210Z   triton_flex_attention_986 0.0160 ms 59.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6140854Z   triton_flex_attention_978 0.0168 ms 57.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6140982Z SingleProcess AUTOTUNE benchmarking takes 0.2178 seconds and 0.3419 seconds precompiling for 24 choices
2025-12-04T09:45:17.6141024Z Autotune Choices Stats:
2025-12-04T09:45:17.6141793Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1007", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017878999933600426, "best_triton_pos": 0}
2025-12-04T09:45:17.6142034Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6142212Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6142486Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6143117Z   triton_flex_attention_backward_1007 0.0179 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6143748Z   triton_flex_attention_backward_1001 0.0207 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6144363Z   triton_flex_attention_backward_998 0.0216 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6144993Z   triton_flex_attention_backward_999 0.0217 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6145618Z   triton_flex_attention_backward_1009 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6146257Z   triton_flex_attention_backward_1008 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6146891Z   triton_flex_attention_backward_1006 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6147518Z   triton_flex_attention_backward_1011 0.0253 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6148141Z   triton_flex_attention_backward_1002 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6148755Z   triton_flex_attention_backward_993 0.0264 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6148882Z SingleProcess AUTOTUNE benchmarking takes 0.2463 seconds and 0.8596 seconds precompiling for 22 choices
2025-12-04T09:45:17.6148956Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6148999Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6149038Z unimplemented []
2025-12-04T09:45:17.6149099Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6149210Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6149786Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6149824Z graph_break []
2025-12-04T09:45:17.6149907Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6149948Z Autotune Choices Stats:
2025-12-04T09:45:17.6150737Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1018", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.6150877Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6150991Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6151153Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6151768Z   triton_flex_attention_1018 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6152367Z   triton_flex_attention_1016 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6152971Z   triton_flex_attention_1019 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6153583Z   triton_flex_attention_1014 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6154188Z   triton_flex_attention_1017 0.0126 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6154806Z   triton_flex_attention_1015 0.0142 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6155437Z   triton_flex_attention_1034 0.0149 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6156042Z   triton_flex_attention_1026 0.0153 ms 65.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6156643Z   triton_flex_attention_1032 0.0163 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6157237Z   triton_flex_attention_1024 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6157363Z SingleProcess AUTOTUNE benchmarking takes 0.2067 seconds and 0.4265 seconds precompiling for 24 choices
2025-12-04T09:45:17.6157404Z Autotune Choices Stats:
2025-12-04T09:45:17.6158179Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1053", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.6158399Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6158572Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6158854Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6159491Z   triton_flex_attention_backward_1053 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6160113Z   triton_flex_attention_backward_1047 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6160763Z   triton_flex_attention_backward_1045 0.0215 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6161383Z   triton_flex_attention_backward_1044 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6162026Z   triton_flex_attention_backward_1055 0.0231 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6162649Z   triton_flex_attention_backward_1054 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6163283Z   triton_flex_attention_backward_1052 0.0250 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6163927Z   triton_flex_attention_backward_1057 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6164552Z   triton_flex_attention_backward_1048 0.0263 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6165175Z   triton_flex_attention_backward_1039 0.0265 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6165304Z SingleProcess AUTOTUNE benchmarking takes 0.2471 seconds and 0.8682 seconds precompiling for 22 choices
2025-12-04T09:45:17.6165380Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6165421Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6165461Z unimplemented []
2025-12-04T09:45:17.6165521Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6165622Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6166203Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6166243Z graph_break []
2025-12-04T09:45:17.6166316Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6166357Z Autotune Choices Stats:
2025-12-04T09:45:17.6167096Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1064", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009960000403225422, "best_triton_pos": 0}
2025-12-04T09:45:17.6167241Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6167356Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6167526Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6168132Z   triton_flex_attention_1064 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6168737Z   triton_flex_attention_1062 0.0109 ms 91.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6169345Z   triton_flex_attention_1065 0.0111 ms 89.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6169956Z   triton_flex_attention_1060 0.0121 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6170608Z   triton_flex_attention_1063 0.0124 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6171210Z   triton_flex_attention_1061 0.0143 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6171843Z   triton_flex_attention_1080 0.0145 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6172455Z   triton_flex_attention_1072 0.0150 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6173054Z   triton_flex_attention_1078 0.0159 ms 62.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6173664Z   triton_flex_attention_1070 0.0166 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6173793Z SingleProcess AUTOTUNE benchmarking takes 0.2080 seconds and 0.3697 seconds precompiling for 24 choices
2025-12-04T09:45:17.6173834Z Autotune Choices Stats:
2025-12-04T09:45:17.6174590Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1099", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.6174815Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6174981Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6175253Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6175888Z   triton_flex_attention_backward_1099 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6176529Z   triton_flex_attention_backward_1093 0.0213 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6177150Z   triton_flex_attention_backward_1090 0.0215 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6177772Z   triton_flex_attention_backward_1091 0.0216 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6178392Z   triton_flex_attention_backward_1101 0.0231 ms 78.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6179026Z   triton_flex_attention_backward_1100 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6179646Z   triton_flex_attention_backward_1098 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6180285Z   triton_flex_attention_backward_1103 0.0253 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6180954Z   triton_flex_attention_backward_1094 0.0260 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6181576Z   triton_flex_attention_backward_1085 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6181705Z SingleProcess AUTOTUNE benchmarking takes 0.2488 seconds and 0.6672 seconds precompiling for 22 choices
2025-12-04T09:45:17.6181778Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6181820Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6181859Z unimplemented []
2025-12-04T09:45:17.6181920Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6182019Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6182592Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6182630Z graph_break []
2025-12-04T09:45:17.6182703Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6182742Z Autotune Choices Stats:
2025-12-04T09:45:17.6183484Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1110", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.6183611Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6183739Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6183916Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6184561Z   triton_flex_attention_1110 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6185157Z   triton_flex_attention_1111 0.0111 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6185759Z   triton_flex_attention_1108 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6186368Z   triton_flex_attention_1109 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6186978Z   triton_flex_attention_1106 0.0126 ms 80.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6187590Z   triton_flex_attention_1107 0.0145 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6188196Z   triton_flex_attention_1126 0.0147 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6188824Z   triton_flex_attention_1118 0.0153 ms 66.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6189436Z   triton_flex_attention_1124 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6190036Z   triton_flex_attention_1116 0.0168 ms 60.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6190169Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.3549 seconds precompiling for 24 choices
2025-12-04T09:45:17.6190210Z Autotune Choices Stats:
2025-12-04T09:45:17.6190997Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1145", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.6191214Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6191381Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6191677Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6192305Z   triton_flex_attention_backward_1145 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6192952Z   triton_flex_attention_backward_1139 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6193593Z   triton_flex_attention_backward_1136 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6194215Z   triton_flex_attention_backward_1137 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6194838Z   triton_flex_attention_backward_1146 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6195461Z   triton_flex_attention_backward_1147 0.0232 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6196102Z   triton_flex_attention_backward_1144 0.0252 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6196726Z   triton_flex_attention_backward_1149 0.0253 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6197379Z   triton_flex_attention_backward_1140 0.0263 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6197998Z   triton_flex_attention_backward_1131 0.0266 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6198128Z SingleProcess AUTOTUNE benchmarking takes 0.2541 seconds and 0.6797 seconds precompiling for 22 choices
2025-12-04T09:45:17.6198202Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6198245Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6198287Z unimplemented []
2025-12-04T09:45:17.6198348Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6198448Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6199021Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6199060Z graph_break []
2025-12-04T09:45:17.6199132Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6199173Z Autotune Choices Stats:
2025-12-04T09:45:17.6199934Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1156", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010559000074863434, "best_triton_pos": 0}
2025-12-04T09:45:17.6200061Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6200177Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6200338Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6200982Z   triton_flex_attention_1156 0.0106 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6201627Z   triton_flex_attention_1154 0.0114 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6202231Z   triton_flex_attention_1157 0.0115 ms 91.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6202840Z   triton_flex_attention_1152 0.0122 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6203446Z   triton_flex_attention_1155 0.0127 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6204052Z   triton_flex_attention_1153 0.0143 ms 73.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6204676Z   triton_flex_attention_1172 0.0145 ms 72.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6205277Z   triton_flex_attention_1164 0.0151 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6205910Z   triton_flex_attention_1170 0.0160 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6206508Z   triton_flex_attention_1150 0.0166 ms 63.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6206639Z SingleProcess AUTOTUNE benchmarking takes 0.2092 seconds and 0.3282 seconds precompiling for 24 choices
2025-12-04T09:45:17.6206681Z Autotune Choices Stats:
2025-12-04T09:45:17.6207449Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1191", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017798999324440956, "best_triton_pos": 0}
2025-12-04T09:45:17.6207667Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6207833Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6208112Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6208751Z   triton_flex_attention_backward_1191 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6209376Z   triton_flex_attention_backward_1185 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6210025Z   triton_flex_attention_backward_1182 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6210676Z   triton_flex_attention_backward_1183 0.0217 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6211305Z   triton_flex_attention_backward_1192 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6211937Z   triton_flex_attention_backward_1193 0.0232 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6212558Z   triton_flex_attention_backward_1190 0.0249 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6213197Z   triton_flex_attention_backward_1195 0.0253 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6213822Z   triton_flex_attention_backward_1186 0.0262 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6214475Z   triton_flex_attention_backward_1177 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6214606Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.6703 seconds precompiling for 22 choices
2025-12-04T09:45:17.6214680Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6214724Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6214762Z unimplemented []
2025-12-04T09:45:17.6214825Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6214924Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6215503Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.6215542Z graph_break []
2025-12-04T09:45:17.6215617Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6215658Z Autotune Choices Stats:
2025-12-04T09:45:17.6216391Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1202", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.6216519Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6216631Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6216794Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6217419Z   triton_flex_attention_1202 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6218024Z   triton_flex_attention_1200 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6218658Z   triton_flex_attention_1203 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6219258Z   triton_flex_attention_1198 0.0124 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6219849Z   triton_flex_attention_1201 0.0126 ms 80.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6220479Z   triton_flex_attention_1199 0.0146 ms 69.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6221083Z   triton_flex_attention_1218 0.0149 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6221709Z   triton_flex_attention_1210 0.0154 ms 65.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6222311Z   triton_flex_attention_1216 0.0164 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6222950Z   triton_flex_attention_1196 0.0169 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6223081Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.5746 seconds precompiling for 24 choices
2025-12-04T09:45:17.6223121Z Autotune Choices Stats:
2025-12-04T09:45:17.6223882Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1237", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.6224100Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6224268Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6224548Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6225183Z   triton_flex_attention_backward_1237 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6225822Z   triton_flex_attention_backward_1231 0.0212 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6226444Z   triton_flex_attention_backward_1228 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6227093Z   triton_flex_attention_backward_1229 0.0217 ms 82.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6227717Z   triton_flex_attention_backward_1239 0.0233 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6228349Z   triton_flex_attention_backward_1238 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6228996Z   triton_flex_attention_backward_1241 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6229639Z   triton_flex_attention_backward_1236 0.0255 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6230298Z   triton_flex_attention_backward_1232 0.0264 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6230948Z   triton_flex_attention_backward_1223 0.0264 ms 68.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6231100Z SingleProcess AUTOTUNE benchmarking takes 0.2354 seconds and 0.7927 seconds precompiling for 22 choices
2025-12-04T09:45:17.6231177Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6231220Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6231271Z unimplemented []
2025-12-04T09:45:17.6231333Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6231434Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6232013Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.6232053Z graph_break []
2025-12-04T09:45:17.6232127Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6232169Z Autotune Choices Stats:
2025-12-04T09:45:17.6232901Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1248", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010080000385642052, "best_triton_pos": 0}
2025-12-04T09:45:17.6233028Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6233142Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6233303Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6233948Z   triton_flex_attention_1248 0.0101 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6234557Z   triton_flex_attention_1246 0.0110 ms 91.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6235170Z   triton_flex_attention_1249 0.0116 ms 87.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6235788Z   triton_flex_attention_1247 0.0122 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6236385Z   triton_flex_attention_1244 0.0124 ms 81.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6236986Z   triton_flex_attention_1245 0.0142 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6237582Z   triton_flex_attention_1264 0.0148 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6238193Z   triton_flex_attention_1256 0.0151 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6238797Z   triton_flex_attention_1262 0.0160 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6239404Z   triton_flex_attention_1242 0.0166 ms 60.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6239551Z SingleProcess AUTOTUNE benchmarking takes 0.2098 seconds and 0.3634 seconds precompiling for 24 choices
2025-12-04T09:45:17.6239592Z Autotune Choices Stats:
2025-12-04T09:45:17.6240368Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1283", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018038999289274216, "best_triton_pos": 0}
2025-12-04T09:45:17.6240622Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6240789Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6241065Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6241696Z   triton_flex_attention_backward_1283 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6242339Z   triton_flex_attention_backward_1277 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6242994Z   triton_flex_attention_backward_1274 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6243615Z   triton_flex_attention_backward_1275 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6244271Z   triton_flex_attention_backward_1285 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6244898Z   triton_flex_attention_backward_1284 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6245522Z   triton_flex_attention_backward_1287 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6246163Z   triton_flex_attention_backward_1282 0.0253 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6246798Z   triton_flex_attention_backward_1278 0.0262 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6247422Z   triton_flex_attention_backward_1269 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6247565Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8755 seconds precompiling for 22 choices
2025-12-04T09:45:17.6247638Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6247690Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6247728Z unimplemented []
2025-12-04T09:45:17.6247790Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6247889Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6248482Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6248521Z graph_break []
2025-12-04T09:45:17.6248594Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6248636Z Autotune Choices Stats:
2025-12-04T09:45:17.6249366Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1294", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.6249494Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6249608Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6249772Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6250389Z   triton_flex_attention_1294 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6251042Z   triton_flex_attention_1292 0.0110 ms 93.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6251650Z   triton_flex_attention_1295 0.0118 ms 86.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6252261Z   triton_flex_attention_1290 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6252878Z   triton_flex_attention_1293 0.0126 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6253479Z   triton_flex_attention_1291 0.0143 ms 72.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6254084Z   triton_flex_attention_1310 0.0148 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6254692Z   triton_flex_attention_1302 0.0153 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6255304Z   triton_flex_attention_1308 0.0162 ms 63.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6255907Z   triton_flex_attention_1288 0.0169 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6256045Z SingleProcess AUTOTUNE benchmarking takes 0.2095 seconds and 0.3664 seconds precompiling for 24 choices
2025-12-04T09:45:17.6256084Z Autotune Choices Stats:
2025-12-04T09:45:17.6256870Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1329", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01823900081217289, "best_triton_pos": 0}
2025-12-04T09:45:17.6257086Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6257252Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6257533Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6258163Z   triton_flex_attention_backward_1329 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6258787Z   triton_flex_attention_backward_1323 0.0210 ms 86.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6259419Z   triton_flex_attention_backward_1321 0.0215 ms 84.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6260042Z   triton_flex_attention_backward_1320 0.0216 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6260714Z   triton_flex_attention_backward_1331 0.0232 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6261363Z   triton_flex_attention_backward_1330 0.0232 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6261988Z   triton_flex_attention_backward_1333 0.0251 ms 72.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6262611Z   triton_flex_attention_backward_1328 0.0253 ms 72.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6263237Z   triton_flex_attention_backward_1324 0.0260 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6263872Z   triton_flex_attention_backward_1315 0.0266 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6264004Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.8094 seconds precompiling for 22 choices
2025-12-04T09:45:17.6264082Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6264128Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6264173Z unimplemented []
2025-12-04T09:45:17.6264233Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6264339Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6264930Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.6264986Z graph_break []
2025-12-04T09:45:17.6265059Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6265100Z Autotune Choices Stats:
2025-12-04T09:45:17.6265851Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1340", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009839000180363655, "best_triton_pos": 0}
2025-12-04T09:45:17.6265979Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6266096Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6266257Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6266870Z   triton_flex_attention_1340 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6267501Z   triton_flex_attention_1341 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6268116Z   triton_flex_attention_1338 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6268716Z   triton_flex_attention_1336 0.0125 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6269338Z   triton_flex_attention_1339 0.0127 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6269947Z   triton_flex_attention_1337 0.0144 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6270582Z   triton_flex_attention_1356 0.0145 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6271194Z   triton_flex_attention_1348 0.0151 ms 65.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6271805Z   triton_flex_attention_1354 0.0161 ms 61.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6272448Z   triton_flex_attention_1346 0.0166 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6272577Z SingleProcess AUTOTUNE benchmarking takes 0.2304 seconds and 0.4372 seconds precompiling for 24 choices
2025-12-04T09:45:17.6272626Z Autotune Choices Stats:
2025-12-04T09:45:17.6273388Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1375", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0176790002733469, "best_triton_pos": 0}
2025-12-04T09:45:17.6273635Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6273829Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6274107Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6274740Z   triton_flex_attention_backward_1375 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6275367Z   triton_flex_attention_backward_1369 0.0209 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6276015Z   triton_flex_attention_backward_1366 0.0215 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6276646Z   triton_flex_attention_backward_1367 0.0216 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6277278Z   triton_flex_attention_backward_1377 0.0231 ms 76.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6277911Z   triton_flex_attention_backward_1376 0.0234 ms 75.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6278555Z   triton_flex_attention_backward_1374 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6279180Z   triton_flex_attention_backward_1379 0.0254 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6279807Z   triton_flex_attention_backward_1361 0.0261 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6280448Z   triton_flex_attention_backward_1370 0.0262 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6280577Z SingleProcess AUTOTUNE benchmarking takes 0.2454 seconds and 0.7164 seconds precompiling for 22 choices
2025-12-04T09:45:17.6280649Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6280693Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6280732Z unimplemented []
2025-12-04T09:45:17.6280808Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6280908Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6281480Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6281537Z graph_break []
2025-12-04T09:45:17.6281611Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6281664Z Autotune Choices Stats:
2025-12-04T09:45:17.6282406Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1386", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01015899982303381, "best_triton_pos": 0}
2025-12-04T09:45:17.6282534Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6282648Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6282811Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6283426Z   triton_flex_attention_1386 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6284043Z   triton_flex_attention_1384 0.0112 ms 91.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6284661Z   triton_flex_attention_1387 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6285283Z   triton_flex_attention_1385 0.0123 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6285882Z   triton_flex_attention_1382 0.0125 ms 81.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6286508Z   triton_flex_attention_1383 0.0143 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6287115Z   triton_flex_attention_1402 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6287711Z   triton_flex_attention_1394 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6288313Z   triton_flex_attention_1400 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6288916Z   triton_flex_attention_1380 0.0166 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6289044Z SingleProcess AUTOTUNE benchmarking takes 0.2108 seconds and 0.3546 seconds precompiling for 24 choices
2025-12-04T09:45:17.6289086Z Autotune Choices Stats:
2025-12-04T09:45:17.6289861Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1421", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.6290076Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6290252Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6290633Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6291279Z   triton_flex_attention_backward_1421 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6291907Z   triton_flex_attention_backward_1415 0.0212 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6292534Z   triton_flex_attention_backward_1413 0.0215 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6293175Z   triton_flex_attention_backward_1412 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6293833Z   triton_flex_attention_backward_1423 0.0233 ms 77.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6294460Z   triton_flex_attention_backward_1422 0.0234 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6295107Z   triton_flex_attention_backward_1420 0.0254 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6295761Z   triton_flex_attention_backward_1425 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6296386Z   triton_flex_attention_backward_1407 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6297007Z   triton_flex_attention_backward_1416 0.0266 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6297137Z SingleProcess AUTOTUNE benchmarking takes 0.2495 seconds and 0.6825 seconds precompiling for 22 choices
2025-12-04T09:45:17.6297212Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6297255Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6297300Z unimplemented []
2025-12-04T09:45:17.6297360Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6297461Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6298049Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.6298089Z graph_break []
2025-12-04T09:45:17.6298162Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6298204Z Autotune Choices Stats:
2025-12-04T09:45:17.6298947Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1432", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.6299095Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6299226Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6299387Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6300003Z   triton_flex_attention_1432 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6300642Z   triton_flex_attention_1430 0.0109 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6301253Z   triton_flex_attention_1433 0.0111 ms 89.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6301864Z   triton_flex_attention_1431 0.0123 ms 80.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6302496Z   triton_flex_attention_1428 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6303098Z   triton_flex_attention_1429 0.0144 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6303735Z   triton_flex_attention_1448 0.0146 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6304370Z   triton_flex_attention_1440 0.0151 ms 65.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6305058Z   triton_flex_attention_1446 0.0159 ms 62.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6305759Z   triton_flex_attention_1438 0.0166 ms 59.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6305910Z SingleProcess AUTOTUNE benchmarking takes 0.2194 seconds and 0.3418 seconds precompiling for 24 choices
2025-12-04T09:45:17.6305998Z Autotune Choices Stats:
2025-12-04T09:45:17.6306840Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1467", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017960000783205032, "best_triton_pos": 0}
2025-12-04T09:45:17.6307073Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6307241Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6307527Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6308184Z   triton_flex_attention_backward_1467 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6308808Z   triton_flex_attention_backward_1461 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6309436Z   triton_flex_attention_backward_1459 0.0213 ms 84.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6310055Z   triton_flex_attention_backward_1458 0.0215 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6310734Z   triton_flex_attention_backward_1469 0.0231 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6311372Z   triton_flex_attention_backward_1468 0.0234 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6311998Z   triton_flex_attention_backward_1471 0.0251 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6312662Z   triton_flex_attention_backward_1466 0.0252 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6313287Z   triton_flex_attention_backward_1462 0.0260 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6313909Z   triton_flex_attention_backward_1453 0.0266 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6314039Z SingleProcess AUTOTUNE benchmarking takes 0.2467 seconds and 0.8049 seconds precompiling for 22 choices
2025-12-04T09:45:17.6314116Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6314159Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6314199Z unimplemented []
2025-12-04T09:45:17.6314261Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6314360Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6314924Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 70), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 24), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 8), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6314963Z graph_break []
2025-12-04T09:45:17.6315036Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6315078Z Autotune Choices Stats:
2025-12-04T09:45:17.6315816Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1478", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01003899984061718, "best_triton_pos": 0}
2025-12-04T09:45:17.6315952Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6316065Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6316240Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6316858Z   triton_flex_attention_1478 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6317459Z   triton_flex_attention_1476 0.0108 ms 93.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6318061Z   triton_flex_attention_1479 0.0116 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6318656Z   triton_flex_attention_1474 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6319255Z   triton_flex_attention_1477 0.0124 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6319887Z   triton_flex_attention_1475 0.0147 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6320525Z   triton_flex_attention_1494 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6321163Z   triton_flex_attention_1486 0.0154 ms 65.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6321764Z   triton_flex_attention_1492 0.0159 ms 63.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6322365Z   triton_flex_attention_1472 0.0166 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6322495Z SingleProcess AUTOTUNE benchmarking takes 0.2177 seconds and 0.3850 seconds precompiling for 24 choices
2025-12-04T09:45:17.6322537Z Autotune Choices Stats:
2025-12-04T09:45:17.6323305Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1513", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.6323520Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6323706Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6323983Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6324614Z   triton_flex_attention_backward_1513 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6325270Z   triton_flex_attention_backward_1507 0.0209 ms 86.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6325889Z   triton_flex_attention_backward_1505 0.0214 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6326528Z   triton_flex_attention_backward_1504 0.0216 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6327175Z   triton_flex_attention_backward_1514 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6327825Z   triton_flex_attention_backward_1515 0.0233 ms 77.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6328454Z   triton_flex_attention_backward_1512 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6329078Z   triton_flex_attention_backward_1517 0.0253 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6329738Z   triton_flex_attention_backward_1508 0.0262 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6330361Z   triton_flex_attention_backward_1499 0.0265 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6330521Z SingleProcess AUTOTUNE benchmarking takes 0.2461 seconds and 0.7066 seconds precompiling for 22 choices
2025-12-04T09:45:17.6330597Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6330639Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6330678Z unimplemented []
2025-12-04T09:45:17.6330739Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6330840Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6331411Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.6331450Z graph_break []
2025-12-04T09:45:17.6331525Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6331565Z Autotune Choices Stats:
2025-12-04T09:45:17.6332321Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1524", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.0106800002977252, "best_triton_pos": 0}
2025-12-04T09:45:17.6332450Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6332565Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6332729Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6333351Z   triton_flex_attention_1524 0.0107 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6333981Z   triton_flex_attention_1522 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6334585Z   triton_flex_attention_1525 0.0114 ms 94.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6335183Z   triton_flex_attention_1520 0.0122 ms 87.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6335783Z   triton_flex_attention_1523 0.0124 ms 86.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6336404Z   triton_flex_attention_1521 0.0146 ms 73.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6337014Z   triton_flex_attention_1532 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6337615Z   triton_flex_attention_1540 0.0150 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6338244Z   triton_flex_attention_1538 0.0161 ms 66.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6338848Z   triton_flex_attention_1530 0.0168 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6338977Z SingleProcess AUTOTUNE benchmarking takes 0.2111 seconds and 0.4119 seconds precompiling for 24 choices
2025-12-04T09:45:17.6339017Z Autotune Choices Stats:
2025-12-04T09:45:17.6339772Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1559", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017839999869465828, "best_triton_pos": 0}
2025-12-04T09:45:17.6339990Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6340156Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6340465Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6341109Z   triton_flex_attention_backward_1559 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6341725Z   triton_flex_attention_backward_1553 0.0209 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6342377Z   triton_flex_attention_backward_1551 0.0213 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6342997Z   triton_flex_attention_backward_1550 0.0214 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6343621Z   triton_flex_attention_backward_1561 0.0230 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6344243Z   triton_flex_attention_backward_1560 0.0231 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6344867Z   triton_flex_attention_backward_1558 0.0250 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6345505Z   triton_flex_attention_backward_1563 0.0251 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6346125Z   triton_flex_attention_backward_1554 0.0260 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6346775Z   triton_flex_attention_backward_1545 0.0263 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6346905Z SingleProcess AUTOTUNE benchmarking takes 0.2489 seconds and 0.8015 seconds precompiling for 22 choices
2025-12-04T09:45:17.6346978Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6347021Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6347060Z unimplemented []
2025-12-04T09:45:17.6347122Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6347220Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6347796Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6347835Z graph_break []
2025-12-04T09:45:17.6347907Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6347948Z Autotune Choices Stats:
2025-12-04T09:45:17.6348692Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1570", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.6348819Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6348931Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6349106Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6349722Z   triton_flex_attention_1570 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6350326Z   triton_flex_attention_1571 0.0112 ms 92.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6350978Z   triton_flex_attention_1568 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6351578Z   triton_flex_attention_1566 0.0124 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6352181Z   triton_flex_attention_1569 0.0128 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6352785Z   triton_flex_attention_1567 0.0145 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6353405Z   triton_flex_attention_1586 0.0147 ms 70.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6356780Z   triton_flex_attention_1578 0.0150 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6357421Z   triton_flex_attention_1584 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6358048Z   triton_flex_attention_1576 0.0168 ms 61.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6358183Z SingleProcess AUTOTUNE benchmarking takes 0.2104 seconds and 0.4599 seconds precompiling for 24 choices
2025-12-04T09:45:17.6358228Z Autotune Choices Stats:
2025-12-04T09:45:17.6359000Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1605", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01807899959385395, "best_triton_pos": 0}
2025-12-04T09:45:17.6359223Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6359393Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6359675Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6360317Z   triton_flex_attention_backward_1605 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6360995Z   triton_flex_attention_backward_1599 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6361619Z   triton_flex_attention_backward_1596 0.0213 ms 84.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6362278Z   triton_flex_attention_backward_1597 0.0215 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6362897Z   triton_flex_attention_backward_1607 0.0233 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6363520Z   triton_flex_attention_backward_1606 0.0234 ms 77.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6364143Z   triton_flex_attention_backward_1604 0.0252 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6364791Z   triton_flex_attention_backward_1609 0.0253 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6365416Z   triton_flex_attention_backward_1600 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6366052Z   triton_flex_attention_backward_1591 0.0268 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6366192Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.6867 seconds precompiling for 22 choices
2025-12-04T09:45:17.6366282Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6366328Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6366370Z unimplemented []
2025-12-04T09:45:17.6366434Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6366538Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6367107Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 73), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 27), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 11), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6367147Z graph_break []
2025-12-04T09:45:17.6367223Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6367265Z Autotune Choices Stats:
2025-12-04T09:45:17.6368002Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1616", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009879999794065952, "best_triton_pos": 0}
2025-12-04T09:45:17.6368131Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6368249Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6368409Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6369031Z   triton_flex_attention_1616 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6369635Z   triton_flex_attention_1614 0.0110 ms 90.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6370247Z   triton_flex_attention_1617 0.0115 ms 86.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6370908Z   triton_flex_attention_1612 0.0121 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6371512Z   triton_flex_attention_1615 0.0124 ms 79.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6372108Z   triton_flex_attention_1613 0.0144 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6372717Z   triton_flex_attention_1632 0.0147 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6373322Z   triton_flex_attention_1624 0.0153 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6373922Z   triton_flex_attention_1630 0.0161 ms 61.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6374538Z   triton_flex_attention_1610 0.0165 ms 59.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6374680Z SingleProcess AUTOTUNE benchmarking takes 0.2088 seconds and 0.5041 seconds precompiling for 24 choices
2025-12-04T09:45:17.6374732Z Autotune Choices Stats:
2025-12-04T09:45:17.6375485Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1651", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018118999898433685, "best_triton_pos": 0}
2025-12-04T09:45:17.6375704Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6375872Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6376148Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6376786Z   triton_flex_attention_backward_1651 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6377416Z   triton_flex_attention_backward_1645 0.0210 ms 86.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6378041Z   triton_flex_attention_backward_1643 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6378669Z   triton_flex_attention_backward_1642 0.0216 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6379321Z   triton_flex_attention_backward_1652 0.0232 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6379947Z   triton_flex_attention_backward_1653 0.0233 ms 77.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6380596Z   triton_flex_attention_backward_1650 0.0252 ms 71.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6381220Z   triton_flex_attention_backward_1655 0.0254 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6381856Z   triton_flex_attention_backward_1646 0.0263 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6382473Z   triton_flex_attention_backward_1637 0.0264 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6382615Z SingleProcess AUTOTUNE benchmarking takes 0.2631 seconds and 0.7101 seconds precompiling for 22 choices
2025-12-04T09:45:17.6382703Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6382748Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6382786Z unimplemented []
2025-12-04T09:45:17.6382848Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6382948Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6383527Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6383565Z graph_break []
2025-12-04T09:45:17.6383641Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6383681Z Autotune Choices Stats:
2025-12-04T09:45:17.6384442Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1662", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009999999776482582, "best_triton_pos": 0}
2025-12-04T09:45:17.6384571Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6384686Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6384848Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6385460Z   triton_flex_attention_1662 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6386070Z   triton_flex_attention_1660 0.0107 ms 93.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6386667Z   triton_flex_attention_1663 0.0108 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6387279Z   triton_flex_attention_1658 0.0121 ms 82.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6387888Z   triton_flex_attention_1661 0.0123 ms 81.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6388489Z   triton_flex_attention_1659 0.0145 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6389096Z   triton_flex_attention_1678 0.0148 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6389704Z   triton_flex_attention_1670 0.0152 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6390319Z   triton_flex_attention_1676 0.0162 ms 61.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6390950Z   triton_flex_attention_1656 0.0169 ms 59.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6391095Z SingleProcess AUTOTUNE benchmarking takes 0.1973 seconds and 0.5238 seconds precompiling for 24 choices
2025-12-04T09:45:17.6391148Z Autotune Choices Stats:
2025-12-04T09:45:17.6391910Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1697", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018079999834299088, "best_triton_pos": 0}
2025-12-04T09:45:17.6392127Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6392297Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6392572Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6393198Z   triton_flex_attention_backward_1697 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6393822Z   triton_flex_attention_backward_1691 0.0210 ms 85.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6394478Z   triton_flex_attention_backward_1689 0.0214 ms 84.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6395102Z   triton_flex_attention_backward_1688 0.0216 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6395738Z   triton_flex_attention_backward_1699 0.0230 ms 78.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6396385Z   triton_flex_attention_backward_1698 0.0231 ms 78.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6397013Z   triton_flex_attention_backward_1701 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6397651Z   triton_flex_attention_backward_1696 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6398295Z   triton_flex_attention_backward_1692 0.0262 ms 69.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6398944Z   triton_flex_attention_backward_1683 0.0266 ms 67.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6399074Z SingleProcess AUTOTUNE benchmarking takes 0.2446 seconds and 0.7318 seconds precompiling for 22 choices
2025-12-04T09:45:17.6399150Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6399192Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6399233Z unimplemented []
2025-12-04T09:45:17.6399293Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6399403Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6399979Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6400018Z graph_break []
2025-12-04T09:45:17.6400100Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6400144Z Autotune Choices Stats:
2025-12-04T09:45:17.6400933Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1708", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.6401060Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6401179Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6401340Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6401965Z   triton_flex_attention_1708 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6402562Z   triton_flex_attention_1706 0.0107 ms 93.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6403188Z   triton_flex_attention_1709 0.0110 ms 91.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6403791Z   triton_flex_attention_1704 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6404432Z   triton_flex_attention_1707 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6405032Z   triton_flex_attention_1705 0.0144 ms 69.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6405637Z   triton_flex_attention_1724 0.0146 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6406257Z   triton_flex_attention_1716 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6406857Z   triton_flex_attention_1722 0.0160 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6407471Z   triton_flex_attention_1702 0.0166 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6407600Z SingleProcess AUTOTUNE benchmarking takes 0.1988 seconds and 0.5275 seconds precompiling for 24 choices
2025-12-04T09:45:17.6407643Z Autotune Choices Stats:
2025-12-04T09:45:17.6408394Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1743", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.6408629Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6408803Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6409080Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6409713Z   triton_flex_attention_backward_1743 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6410344Z   triton_flex_attention_backward_1737 0.0208 ms 85.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6411026Z   triton_flex_attention_backward_1734 0.0213 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6411678Z   triton_flex_attention_backward_1735 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6412305Z   triton_flex_attention_backward_1745 0.0232 ms 76.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6412973Z   triton_flex_attention_backward_1744 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6413596Z   triton_flex_attention_backward_1742 0.0249 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6414220Z   triton_flex_attention_backward_1747 0.0252 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6414850Z   triton_flex_attention_backward_1738 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6415471Z   triton_flex_attention_backward_1729 0.0264 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6415600Z SingleProcess AUTOTUNE benchmarking takes 0.2428 seconds and 0.7372 seconds precompiling for 22 choices
2025-12-04T09:45:17.6415675Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6415728Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6415766Z unimplemented []
2025-12-04T09:45:17.6415826Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6415925Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6416491Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6416547Z graph_break []
2025-12-04T09:45:17.6416621Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6416661Z Autotune Choices Stats:
2025-12-04T09:45:17.6417422Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1754", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009878999553620815, "best_triton_pos": 0}
2025-12-04T09:45:17.6417551Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6417667Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6417827Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6418439Z   triton_flex_attention_1754 0.0099 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6419039Z   triton_flex_attention_1752 0.0110 ms 90.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6419648Z   triton_flex_attention_1755 0.0114 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6420263Z   triton_flex_attention_1753 0.0124 ms 79.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6420881Z   triton_flex_attention_1750 0.0125 ms 79.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6421519Z   triton_flex_attention_1751 0.0143 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6422125Z   triton_flex_attention_1770 0.0149 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6422728Z   triton_flex_attention_1762 0.0152 ms 65.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6423330Z   triton_flex_attention_1768 0.0163 ms 60.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6423933Z   triton_flex_attention_1748 0.0170 ms 58.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6424067Z SingleProcess AUTOTUNE benchmarking takes 0.2060 seconds and 0.4503 seconds precompiling for 24 choices
2025-12-04T09:45:17.6424107Z Autotune Choices Stats:
2025-12-04T09:45:17.6424879Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1789", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.6425109Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6425382Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6425677Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6426308Z   triton_flex_attention_backward_1789 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6426946Z   triton_flex_attention_backward_1783 0.0209 ms 87.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6427568Z   triton_flex_attention_backward_1780 0.0216 ms 84.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6428219Z   triton_flex_attention_backward_1781 0.0217 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6428881Z   triton_flex_attention_backward_1791 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6429497Z   triton_flex_attention_backward_1790 0.0235 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6430147Z   triton_flex_attention_backward_1788 0.0254 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6430797Z   triton_flex_attention_backward_1793 0.0255 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6431422Z   triton_flex_attention_backward_1775 0.0264 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6432068Z   triton_flex_attention_backward_1784 0.0265 ms 68.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6432200Z SingleProcess AUTOTUNE benchmarking takes 0.2498 seconds and 0.6949 seconds precompiling for 22 choices
2025-12-04T09:45:17.6432278Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6432321Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6432368Z unimplemented []
2025-12-04T09:45:17.6432433Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6432540Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6433127Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6433172Z graph_break []
2025-12-04T09:45:17.6433250Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6433295Z Autotune Choices Stats:
2025-12-04T09:45:17.6434047Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1800", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.6434201Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6434329Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6434493Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6435119Z   triton_flex_attention_1800 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6435742Z   triton_flex_attention_1798 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6436347Z   triton_flex_attention_1801 0.0115 ms 88.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6436953Z   triton_flex_attention_1796 0.0121 ms 83.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6437565Z   triton_flex_attention_1799 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6438173Z   triton_flex_attention_1816 0.0145 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6438798Z   triton_flex_attention_1797 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6439405Z   triton_flex_attention_1808 0.0152 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6440035Z   triton_flex_attention_1814 0.0161 ms 63.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6440649Z   triton_flex_attention_1806 0.0168 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6440780Z SingleProcess AUTOTUNE benchmarking takes 0.2107 seconds and 0.5450 seconds precompiling for 24 choices
2025-12-04T09:45:17.6440821Z Autotune Choices Stats:
2025-12-04T09:45:17.6441595Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1835", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017999999225139618, "best_triton_pos": 0}
2025-12-04T09:45:17.6441812Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6441979Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6442267Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6442934Z   triton_flex_attention_backward_1835 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6443559Z   triton_flex_attention_backward_1829 0.0210 ms 85.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6444186Z   triton_flex_attention_backward_1826 0.0212 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6444815Z   triton_flex_attention_backward_1827 0.0213 ms 84.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6445443Z   triton_flex_attention_backward_1837 0.0231 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6446084Z   triton_flex_attention_backward_1836 0.0232 ms 77.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6446702Z   triton_flex_attention_backward_1839 0.0252 ms 71.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6447349Z   triton_flex_attention_backward_1834 0.0252 ms 71.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6447974Z   triton_flex_attention_backward_1830 0.0260 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6448602Z   triton_flex_attention_backward_1821 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6448733Z SingleProcess AUTOTUNE benchmarking takes 0.2508 seconds and 0.7770 seconds precompiling for 22 choices
2025-12-04T09:45:17.6448807Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6448850Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6448887Z unimplemented []
2025-12-04T09:45:17.6448948Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6449046Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6449616Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6449654Z graph_break []
2025-12-04T09:45:17.6449730Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6449779Z Autotune Choices Stats:
2025-12-04T09:45:17.6450536Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1846", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010320000350475311, "best_triton_pos": 0}
2025-12-04T09:45:17.6450679Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6450807Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6450967Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6451590Z   triton_flex_attention_1846 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6452192Z   triton_flex_attention_1844 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6452818Z   triton_flex_attention_1847 0.0114 ms 90.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6453422Z   triton_flex_attention_1842 0.0122 ms 84.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6454035Z   triton_flex_attention_1845 0.0124 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6454638Z   triton_flex_attention_1843 0.0144 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6455244Z   triton_flex_attention_1862 0.0146 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6455872Z   triton_flex_attention_1854 0.0154 ms 67.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6456474Z   triton_flex_attention_1860 0.0160 ms 64.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6457071Z   triton_flex_attention_1840 0.0167 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6457201Z SingleProcess AUTOTUNE benchmarking takes 0.2278 seconds and 0.3492 seconds precompiling for 24 choices
2025-12-04T09:45:17.6457243Z Autotune Choices Stats:
2025-12-04T09:45:17.6458003Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1881", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.6458219Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6458395Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6458674Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6459307Z   triton_flex_attention_backward_1881 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6459960Z   triton_flex_attention_backward_1875 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6460611Z   triton_flex_attention_backward_1873 0.0216 ms 83.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6461244Z   triton_flex_attention_backward_1872 0.0216 ms 83.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6461866Z   triton_flex_attention_backward_1882 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6462498Z   triton_flex_attention_backward_1883 0.0231 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6463140Z   triton_flex_attention_backward_1880 0.0254 ms 71.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6463766Z   triton_flex_attention_backward_1885 0.0254 ms 71.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6464425Z   triton_flex_attention_backward_1876 0.0263 ms 68.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6465044Z   triton_flex_attention_backward_1867 0.0267 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6465174Z SingleProcess AUTOTUNE benchmarking takes 0.2409 seconds and 0.8665 seconds precompiling for 22 choices
2025-12-04T09:45:17.6465249Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6465292Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6465331Z unimplemented []
2025-12-04T09:45:17.6465390Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6465490Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6466059Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 74), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 28), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 12), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6466097Z graph_break []
2025-12-04T09:45:17.6466172Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6466213Z Autotune Choices Stats:
2025-12-04T09:45:17.6466972Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1892", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.6467100Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6467215Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6467374Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6467999Z   triton_flex_attention_1892 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6468623Z   triton_flex_attention_1890 0.0109 ms 92.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6469227Z   triton_flex_attention_1893 0.0114 ms 88.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6469824Z   triton_flex_attention_1888 0.0122 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6470465Z   triton_flex_attention_1891 0.0123 ms 81.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6471080Z   triton_flex_attention_1889 0.0144 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6471684Z   triton_flex_attention_1908 0.0146 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6472292Z   triton_flex_attention_1900 0.0151 ms 66.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6472918Z   triton_flex_attention_1906 0.0161 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6473519Z   triton_flex_attention_1886 0.0167 ms 60.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6473648Z SingleProcess AUTOTUNE benchmarking takes 0.2106 seconds and 0.3466 seconds precompiling for 24 choices
2025-12-04T09:45:17.6473688Z Autotune Choices Stats:
2025-12-04T09:45:17.6474442Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1927", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.6474662Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6474828Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6475104Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6475750Z   triton_flex_attention_backward_1927 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6476375Z   triton_flex_attention_backward_1921 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6477032Z   triton_flex_attention_backward_1918 0.0216 ms 82.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6477653Z   triton_flex_attention_backward_1919 0.0216 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6478276Z   triton_flex_attention_backward_1929 0.0231 ms 76.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6478925Z   triton_flex_attention_backward_1928 0.0233 ms 76.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6479560Z   triton_flex_attention_backward_1926 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6480186Z   triton_flex_attention_backward_1931 0.0254 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6480846Z   triton_flex_attention_backward_1922 0.0261 ms 68.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6481496Z   triton_flex_attention_backward_1913 0.0263 ms 67.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6481626Z SingleProcess AUTOTUNE benchmarking takes 0.2431 seconds and 0.7860 seconds precompiling for 22 choices
2025-12-04T09:45:17.6481701Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6481744Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6481781Z unimplemented []
2025-12-04T09:45:17.6481842Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6481939Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6482516Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6482554Z graph_break []
2025-12-04T09:45:17.6482630Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6482671Z Autotune Choices Stats:
2025-12-04T09:45:17.6483416Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1938", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.6483545Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6483671Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6483831Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6484435Z   triton_flex_attention_1938 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6485058Z   triton_flex_attention_1936 0.0109 ms 93.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6485688Z   triton_flex_attention_1939 0.0116 ms 88.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6486292Z   triton_flex_attention_1934 0.0122 ms 83.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6486894Z   triton_flex_attention_1937 0.0124 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6487491Z   triton_flex_attention_1935 0.0144 ms 70.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6488110Z   triton_flex_attention_1954 0.0148 ms 68.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6488712Z   triton_flex_attention_1946 0.0154 ms 66.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6489323Z   triton_flex_attention_1952 0.0164 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6489947Z   triton_flex_attention_1944 0.0170 ms 59.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6490079Z SingleProcess AUTOTUNE benchmarking takes 0.2077 seconds and 0.3245 seconds precompiling for 24 choices
2025-12-04T09:45:17.6490120Z Autotune Choices Stats:
2025-12-04T09:45:17.6490908Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_1973", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01783899962902069, "best_triton_pos": 0}
2025-12-04T09:45:17.6491125Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6491291Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6491570Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6492224Z   triton_flex_attention_backward_1973 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6492843Z   triton_flex_attention_backward_1967 0.0211 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6493475Z   triton_flex_attention_backward_1965 0.0216 ms 82.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6494120Z   triton_flex_attention_backward_1964 0.0217 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6494753Z   triton_flex_attention_backward_1975 0.0233 ms 76.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6495393Z   triton_flex_attention_backward_1974 0.0235 ms 76.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6496026Z   triton_flex_attention_backward_1972 0.0253 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6496672Z   triton_flex_attention_backward_1977 0.0255 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6497298Z   triton_flex_attention_backward_1968 0.0266 ms 67.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6497926Z   triton_flex_attention_backward_1959 0.0266 ms 67.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6498076Z SingleProcess AUTOTUNE benchmarking takes 0.2453 seconds and 0.8096 seconds precompiling for 22 choices
2025-12-04T09:45:17.6498150Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6498193Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6498231Z unimplemented []
2025-12-04T09:45:17.6498291Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6498391Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6498959Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.6498998Z graph_break []
2025-12-04T09:45:17.6499071Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6499113Z Autotune Choices Stats:
2025-12-04T09:45:17.6499857Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_1984", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010440000332891941, "best_triton_pos": 0}
2025-12-04T09:45:17.6499988Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6500103Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6500264Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6500918Z   triton_flex_attention_1984 0.0104 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6501531Z   triton_flex_attention_1982 0.0109 ms 95.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6502169Z   triton_flex_attention_1985 0.0113 ms 92.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6502769Z   triton_flex_attention_1980 0.0122 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6503372Z   triton_flex_attention_1983 0.0124 ms 83.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6503974Z   triton_flex_attention_1981 0.0142 ms 73.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6504579Z   triton_flex_attention_2000 0.0146 ms 71.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6505190Z   triton_flex_attention_1992 0.0151 ms 69.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6505796Z   triton_flex_attention_1998 0.0160 ms 65.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6506420Z   triton_flex_attention_1978 0.0168 ms 62.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6506563Z SingleProcess AUTOTUNE benchmarking takes 0.2059 seconds and 0.3341 seconds precompiling for 24 choices
2025-12-04T09:45:17.6506605Z Autotune Choices Stats:
2025-12-04T09:45:17.6507361Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2019", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018200000748038292, "best_triton_pos": 0}
2025-12-04T09:45:17.6507582Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6507748Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6508025Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6508649Z   triton_flex_attention_backward_2019 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6509287Z   triton_flex_attention_backward_2013 0.0210 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6509908Z   triton_flex_attention_backward_2010 0.0214 ms 85.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6510575Z   triton_flex_attention_backward_2011 0.0214 ms 85.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6511224Z   triton_flex_attention_backward_2021 0.0232 ms 78.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6511850Z   triton_flex_attention_backward_2020 0.0233 ms 78.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6512475Z   triton_flex_attention_backward_2018 0.0250 ms 72.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6513100Z   triton_flex_attention_backward_2023 0.0253 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6513737Z   triton_flex_attention_backward_2014 0.0262 ms 69.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6514358Z   triton_flex_attention_backward_2005 0.0267 ms 68.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6514511Z SingleProcess AUTOTUNE benchmarking takes 0.2422 seconds and 0.7502 seconds precompiling for 22 choices
2025-12-04T09:45:17.6514586Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6514629Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6514668Z unimplemented []
2025-12-04T09:45:17.6514727Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6514826Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6515403Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6515442Z graph_break []
2025-12-04T09:45:17.6515516Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6515556Z Autotune Choices Stats:
2025-12-04T09:45:17.6516296Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2030", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010280000045895576, "best_triton_pos": 0}
2025-12-04T09:45:17.6516424Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6516539Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6516702Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6517311Z   triton_flex_attention_2030 0.0103 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6517924Z   triton_flex_attention_2028 0.0109 ms 94.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6518527Z   triton_flex_attention_2031 0.0112 ms 91.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6519160Z   triton_flex_attention_2026 0.0126 ms 81.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6519761Z   triton_flex_attention_2029 0.0127 ms 80.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6520359Z   triton_flex_attention_2027 0.0142 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6521009Z   triton_flex_attention_2046 0.0147 ms 70.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6521608Z   triton_flex_attention_2038 0.0152 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6522222Z   triton_flex_attention_2044 0.0162 ms 63.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6522816Z   triton_flex_attention_2024 0.0165 ms 62.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6522975Z SingleProcess AUTOTUNE benchmarking takes 0.2047 seconds and 0.3631 seconds precompiling for 24 choices
2025-12-04T09:45:17.6523015Z Autotune Choices Stats:
2025-12-04T09:45:17.6523787Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2065", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017799999564886093, "best_triton_pos": 0}
2025-12-04T09:45:17.6524004Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6524171Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6524448Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6525078Z   triton_flex_attention_backward_2065 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6525721Z   triton_flex_attention_backward_2059 0.0208 ms 85.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6526349Z   triton_flex_attention_backward_2056 0.0213 ms 83.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6526970Z   triton_flex_attention_backward_2057 0.0214 ms 83.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6527626Z   triton_flex_attention_backward_2067 0.0230 ms 77.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6528252Z   triton_flex_attention_backward_2066 0.0234 ms 76.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6528873Z   triton_flex_attention_backward_2064 0.0250 ms 71.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6529496Z   triton_flex_attention_backward_2069 0.0252 ms 70.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6530139Z   triton_flex_attention_backward_2060 0.0260 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6530817Z   triton_flex_attention_backward_2051 0.0263 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6530949Z SingleProcess AUTOTUNE benchmarking takes 0.2494 seconds and 0.8153 seconds precompiling for 22 choices
2025-12-04T09:45:17.6531023Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6531079Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6531117Z unimplemented []
2025-12-04T09:45:17.6531178Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6531289Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6531866Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6531905Z graph_break []
2025-12-04T09:45:17.6531980Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6532020Z Autotune Choices Stats:
2025-12-04T09:45:17.6532764Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2076", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.6532892Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6533007Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6533171Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6533782Z   triton_flex_attention_2076 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6534404Z   triton_flex_attention_2074 0.0108 ms 94.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6535020Z   triton_flex_attention_2077 0.0114 ms 88.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6535620Z   triton_flex_attention_2072 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6536250Z   triton_flex_attention_2075 0.0125 ms 81.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6536856Z   triton_flex_attention_2073 0.0146 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6537459Z   triton_flex_attention_2092 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6538059Z   triton_flex_attention_2084 0.0153 ms 66.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6538669Z   triton_flex_attention_2090 0.0162 ms 62.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6539289Z   triton_flex_attention_2070 0.0167 ms 60.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6539418Z SingleProcess AUTOTUNE benchmarking takes 0.2086 seconds and 0.3462 seconds precompiling for 24 choices
2025-12-04T09:45:17.6539459Z Autotune Choices Stats:
2025-12-04T09:45:17.6540217Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2111", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.017680000513792038, "best_triton_pos": 0}
2025-12-04T09:45:17.6540504Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6540671Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6540951Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6541591Z   triton_flex_attention_backward_2111 0.0177 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6542226Z   triton_flex_attention_backward_2105 0.0210 ms 84.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6542854Z   triton_flex_attention_backward_2102 0.0214 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6543493Z   triton_flex_attention_backward_2103 0.0215 ms 82.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6544119Z   triton_flex_attention_backward_2113 0.0232 ms 76.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6544769Z   triton_flex_attention_backward_2112 0.0234 ms 75.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6545389Z   triton_flex_attention_backward_2110 0.0250 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6546023Z   triton_flex_attention_backward_2115 0.0253 ms 69.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6546661Z   triton_flex_attention_backward_2106 0.0262 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6547284Z   triton_flex_attention_backward_2097 0.0262 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6547423Z SingleProcess AUTOTUNE benchmarking takes 0.2473 seconds and 0.8010 seconds precompiling for 22 choices
2025-12-04T09:45:17.6547499Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6547541Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6547583Z unimplemented []
2025-12-04T09:45:17.6547643Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6547745Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6548317Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 64), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 18), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2), ('coordesc_tuning_bench', 2)]
2025-12-04T09:45:17.6548373Z graph_break []
2025-12-04T09:45:17.6548450Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6548490Z Autotune Choices Stats:
2025-12-04T09:45:17.6549233Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2122", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010040000081062317, "best_triton_pos": 0}
2025-12-04T09:45:17.6549361Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6549475Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6549636Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6550246Z   triton_flex_attention_2122 0.0100 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6550876Z   triton_flex_attention_2120 0.0110 ms 90.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6551479Z   triton_flex_attention_2123 0.0113 ms 88.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6552105Z   triton_flex_attention_2118 0.0122 ms 82.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6552706Z   triton_flex_attention_2121 0.0122 ms 82.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6553343Z   triton_flex_attention_2119 0.0142 ms 70.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6553943Z   triton_flex_attention_2138 0.0147 ms 68.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6554553Z   triton_flex_attention_2130 0.0151 ms 66.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6555164Z   triton_flex_attention_2136 0.0163 ms 61.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6555767Z   triton_flex_attention_2116 0.0167 ms 60.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6555917Z SingleProcess AUTOTUNE benchmarking takes 0.2130 seconds and 0.3464 seconds precompiling for 24 choices
2025-12-04T09:45:17.6555957Z Autotune Choices Stats:
2025-12-04T09:45:17.6556724Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2157", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018039999529719353, "best_triton_pos": 0}
2025-12-04T09:45:17.6556959Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6557125Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6557411Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6558040Z   triton_flex_attention_backward_2157 0.0180 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6558673Z   triton_flex_attention_backward_2151 0.0210 ms 85.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6559302Z   triton_flex_attention_backward_2148 0.0217 ms 83.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6559924Z   triton_flex_attention_backward_2149 0.0217 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6560592Z   triton_flex_attention_backward_2159 0.0234 ms 77.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6561224Z   triton_flex_attention_backward_2158 0.0234 ms 77.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6561878Z   triton_flex_attention_backward_2156 0.0252 ms 71.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6562498Z   triton_flex_attention_backward_2161 0.0256 ms 70.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6563121Z   triton_flex_attention_backward_2152 0.0261 ms 69.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6563764Z   triton_flex_attention_backward_2143 0.0266 ms 67.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6563897Z SingleProcess AUTOTUNE benchmarking takes 0.2464 seconds and 0.8851 seconds precompiling for 22 choices
2025-12-04T09:45:17.6563970Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6564013Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6564050Z unimplemented []
2025-12-04T09:45:17.6564112Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6564212Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6564787Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 65), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 19), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('coordesc_tuning_bench', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6564826Z graph_break []
2025-12-04T09:45:17.6564898Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6564950Z Autotune Choices Stats:
2025-12-04T09:45:17.6565692Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2168", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.009840000420808792, "best_triton_pos": 0}
2025-12-04T09:45:17.6565841Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6565956Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6566118Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6566732Z   triton_flex_attention_2168 0.0098 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6567326Z   triton_flex_attention_2166 0.0108 ms 90.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6567933Z   triton_flex_attention_2169 0.0114 ms 86.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6568544Z   triton_flex_attention_2167 0.0124 ms 79.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6569145Z   triton_flex_attention_2164 0.0124 ms 79.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6569755Z   triton_flex_attention_2165 0.0145 ms 67.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6570386Z   triton_flex_attention_2184 0.0146 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6571032Z   triton_flex_attention_2176 0.0150 ms 65.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6571635Z   triton_flex_attention_2182 0.0160 ms 61.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6572238Z   triton_flex_attention_2174 0.0167 ms 58.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6572370Z SingleProcess AUTOTUNE benchmarking takes 0.2149 seconds and 0.3567 seconds precompiling for 24 choices
2025-12-04T09:45:17.6572411Z Autotune Choices Stats:
2025-12-04T09:45:17.6573211Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2203", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018160000443458557, "best_triton_pos": 0}
2025-12-04T09:45:17.6573430Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6573596Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6573900Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6574551Z   triton_flex_attention_backward_2203 0.0182 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6575173Z   triton_flex_attention_backward_2197 0.0210 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6575812Z   triton_flex_attention_backward_2194 0.0213 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6576456Z   triton_flex_attention_backward_2195 0.0214 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6577085Z   triton_flex_attention_backward_2205 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6577718Z   triton_flex_attention_backward_2204 0.0233 ms 78.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6578338Z   triton_flex_attention_backward_2202 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6578993Z   triton_flex_attention_backward_2207 0.0252 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6579618Z   triton_flex_attention_backward_2198 0.0262 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6580258Z   triton_flex_attention_backward_2189 0.0266 ms 68.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6580388Z SingleProcess AUTOTUNE benchmarking takes 0.2457 seconds and 0.8512 seconds precompiling for 22 choices
2025-12-04T09:45:17.6580607Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6580648Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6580688Z unimplemented []
2025-12-04T09:45:17.6580748Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6580850Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6581424Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 69), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 23), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('coordesc_tuning_bench', 7), ('async_compile_cache_hit', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6581484Z graph_break []
2025-12-04T09:45:17.6581558Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6581598Z Autotune Choices Stats:
2025-12-04T09:45:17.6582342Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2214", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010200000368058681, "best_triton_pos": 0}
2025-12-04T09:45:17.6582498Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6582614Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6582777Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6583428Z   triton_flex_attention_2214 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6584033Z   triton_flex_attention_2212 0.0110 ms 92.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6584635Z   triton_flex_attention_2215 0.0112 ms 91.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6585258Z   triton_flex_attention_2210 0.0123 ms 83.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6585869Z   triton_flex_attention_2213 0.0124 ms 82.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6586471Z   triton_flex_attention_2211 0.0144 ms 70.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6587085Z   triton_flex_attention_2230 0.0148 ms 68.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6587717Z   triton_flex_attention_2222 0.0151 ms 67.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6588326Z   triton_flex_attention_2228 0.0162 ms 62.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6588931Z   triton_flex_attention_2208 0.0167 ms 61.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6589063Z SingleProcess AUTOTUNE benchmarking takes 0.2066 seconds and 0.3920 seconds precompiling for 24 choices
2025-12-04T09:45:17.6589103Z Autotune Choices Stats:
2025-12-04T09:45:17.6589863Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2249", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.018120000138878822, "best_triton_pos": 0}
2025-12-04T09:45:17.6590094Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6590262Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6590576Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6591211Z   triton_flex_attention_backward_2249 0.0181 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6591881Z   triton_flex_attention_backward_2243 0.0210 ms 86.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6592506Z   triton_flex_attention_backward_2241 0.0212 ms 85.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6593125Z   triton_flex_attention_backward_2240 0.0214 ms 84.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6593750Z   triton_flex_attention_backward_2250 0.0230 ms 78.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6594388Z   triton_flex_attention_backward_2251 0.0231 ms 78.5% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6595012Z   triton_flex_attention_backward_2248 0.0251 ms 72.2% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6595651Z   triton_flex_attention_backward_2253 0.0252 ms 71.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6596294Z   triton_flex_attention_backward_2244 0.0261 ms 69.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6596921Z   triton_flex_attention_backward_2235 0.0263 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6597048Z SingleProcess AUTOTUNE benchmarking takes 0.2484 seconds and 0.7948 seconds precompiling for 22 choices
2025-12-04T09:45:17.6597121Z ----------------------------- Captured stdout call -----------------------------
2025-12-04T09:45:17.6597164Z frames [('total', 2), ('ok', 2)]
2025-12-04T09:45:17.6597201Z unimplemented []
2025-12-04T09:45:17.6597262Z stats [('calls_captured', 105), ('unique_graphs', 2)]
2025-12-04T09:45:17.6597361Z aot_autograd [('total', 2), ('autograd_cache_miss', 2), ('autograd_cache_saved', 2), ('ok', 2)]
2025-12-04T09:45:17.6597938Z inductor [('triton_bundler_save_kernel', 528), ('benchmarking.InductorBenchmarker.benchmark_gpu', 68), ('async_compile_cache_miss', 62), ('select_algorithm_num_precompiles', 46), ('benchmarking.InductorBenchmarker.benchmark', 22), ('pattern_matcher_nodes', 16), ('pattern_matcher_count', 14), ('async_compile_cache_hit', 6), ('coordesc_tuning_bench', 6), ('fxgraph_cache_miss', 3), ('triton_bundler_save_static_autotuner', 3), ('select_algorithm_precompile', 2), ('select_algorithm_autotune', 2)]
2025-12-04T09:45:17.6597975Z graph_break []
2025-12-04T09:45:17.6598049Z ----------------------------- Captured stderr call -----------------------------
2025-12-04T09:45:17.6598090Z Autotune Choices Stats:
2025-12-04T09:45:17.6598842Z {"num_choices": 24, "num_triton_choices": 24, "best_kernel": "triton_flex_attention_2260", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.010160000063478947, "best_triton_pos": 0}
2025-12-04T09:45:17.6598975Z AUTOTUNE flex_attention(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6599090Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6599260Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6599893Z   triton_flex_attention_2260 0.0102 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6600533Z   triton_flex_attention_2258 0.0112 ms 90.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6601136Z   triton_flex_attention_2261 0.0114 ms 89.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6601738Z   triton_flex_attention_2256 0.0124 ms 81.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6602340Z   triton_flex_attention_2259 0.0126 ms 80.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6602963Z   triton_flex_attention_2257 0.0142 ms 71.6% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6603571Z   triton_flex_attention_2276 0.0148 ms 68.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6604197Z   triton_flex_attention_2268 0.0152 ms 66.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=64, BLOCK_N=128, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6604835Z   triton_flex_attention_2274 0.0162 ms 62.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=128, BLOCK_N=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6605439Z   triton_flex_attention_2254 0.0168 ms 60.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M=16, BLOCK_N=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, USE_TMA=False, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6605570Z SingleProcess AUTOTUNE benchmarking takes 0.2129 seconds and 0.4452 seconds precompiling for 24 choices
2025-12-04T09:45:17.6605611Z Autotune Choices Stats:
2025-12-04T09:45:17.6606370Z {"num_choices": 22, "num_triton_choices": 22, "best_kernel": "triton_flex_attention_backward_2295", "best_kernel_desc": "BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION=\"'ieee'\", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4", "best_time": 0.01775999926030636, "best_triton_pos": 0}
2025-12-04T09:45:17.6606589Z AUTOTUNE flex_attention_backward(1x2x128x64, 1x2x128x64, 1x2x128x64, 1x2x128, 1x2x128, 1x2x128x64, 1x2x128x64, 1x2x128x64, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1, 1x1x1, 1x1x1x1)
2025-12-04T09:45:17.6606753Z strides: [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [256, 128, 1], [256, 128, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [16384, 8192, 64, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1], [1, 1, 1], [1, 1, 1, 1]
2025-12-04T09:45:17.6607038Z dtypes: torch.float16, torch.float16, torch.float16, torch.float32, torch.float32, torch.float16, torch.float16, torch.float16, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32, torch.int32
2025-12-04T09:45:17.6607672Z   triton_flex_attention_backward_2295 0.0178 ms 100.0% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6608307Z   triton_flex_attention_backward_2289 0.0210 ms 84.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6608945Z   triton_flex_attention_backward_2286 0.0216 ms 82.1% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6609567Z   triton_flex_attention_backward_2287 0.0217 ms 81.8% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=32, BLOCK_N1=32, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6610197Z   triton_flex_attention_backward_2297 0.0232 ms 76.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6610869Z   triton_flex_attention_backward_2296 0.0234 ms 75.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6611506Z   triton_flex_attention_backward_2294 0.0250 ms 70.9% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6612137Z   triton_flex_attention_backward_2299 0.0255 ms 69.7% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=64, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=64, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=8
2025-12-04T09:45:17.6612784Z   triton_flex_attention_backward_2290 0.0264 ms 67.4% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=32, BLOCK_M2=128, BLOCK_N1=128, BLOCK_N2=32, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=0, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6613443Z   triton_flex_attention_backward_2281 0.0264 ms 67.3% BLOCKS_ARE_CONTIGUOUS=False, BLOCK_M1=16, BLOCK_M2=64, BLOCK_N1=64, BLOCK_N2=16, FLOAT32_PRECISION="'ieee'", GQA_SHARED_HEADS=1, HAS_FULL_BLOCKS=True, IS_DIVISIBLE=True, OUTPUT_LOGSUMEXP=True, OUTPUT_MAX=False, PRESCALE_QK=False, QK_HEAD_DIM=64, QK_HEAD_DIM_ROUNDED=64, ROWS_GUARANTEED_SAFE=False, SAFE_HEAD_DIM=True, SM_SCALE=0.125, SPARSE_KV_BLOCK_SIZE=128, SPARSE_Q_BLOCK_SIZE=128, V_HEAD_DIM=64, V_HEAD_DIM_ROUNDED=64, WRITE_DQ=True, kpack=2, matrix_instr_nonkdim=16, waves_per_eu=0, num_stages=1, num_warps=4
2025-12-04T09:45:17.6613573Z SingleProcess AUTOTUNE benchmarking takes 0.2432 seconds and 0.7978 seconds precompiling for 22 choices
2025-12-04T09:45:17.6613808Z - generated xml file: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_flex_attention/inductor.test_flex_attention-f823fef124b8972d.xml -
2025-12-04T09:45:17.6613870Z =========================== short test summary info ============================
2025-12-04T09:45:17.6614180Z FAILED [4.2083s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpzpsl3_e9/flex_attention_configs.json was not created
2025-12-04T09:45:17.6614184Z 
2025-12-04T09:45:17.6614260Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6614426Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6614430Z 
2025-12-04T09:45:17.6614522Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6614792Z FAILED [3.8167s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp_c38ejjw/flex_attention_configs.json was not created
2025-12-04T09:45:17.6614795Z 
2025-12-04T09:45:17.6614869Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6615026Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6615029Z 
2025-12-04T09:45:17.6615116Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6615385Z FAILED [3.6221s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpb8up9rbc/flex_attention_configs.json was not created
2025-12-04T09:45:17.6615388Z 
2025-12-04T09:45:17.6615471Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6615629Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6615632Z 
2025-12-04T09:45:17.6615715Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6615980Z FAILED [3.7765s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpz0pv_51w/flex_attention_configs.json was not created
2025-12-04T09:45:17.6615982Z 
2025-12-04T09:45:17.6616053Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6616220Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6616233Z 
2025-12-04T09:45:17.6616315Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6616579Z FAILED [3.7979s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp7_woya4z/flex_attention_configs.json was not created
2025-12-04T09:45:17.6616581Z 
2025-12-04T09:45:17.6616651Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6616822Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6616824Z 
2025-12-04T09:45:17.6616909Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6617176Z FAILED [3.9433s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmplrzjtur0/flex_attention_configs.json was not created
2025-12-04T09:45:17.6617180Z 
2025-12-04T09:45:17.6617251Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6617408Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6617409Z 
2025-12-04T09:45:17.6617493Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6617759Z FAILED [3.9359s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpdyu0elhy/flex_attention_configs.json was not created
2025-12-04T09:45:17.6617761Z 
2025-12-04T09:45:17.6617832Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6617989Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6617991Z 
2025-12-04T09:45:17.6618074Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6618341Z FAILED [3.7064s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp4919na5_/flex_attention_configs.json was not created
2025-12-04T09:45:17.6618343Z 
2025-12-04T09:45:17.6618412Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6618569Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6618572Z 
2025-12-04T09:45:17.6618654Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6618920Z FAILED [3.6811s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp9gltujgb/flex_attention_configs.json was not created
2025-12-04T09:45:17.6618923Z 
2025-12-04T09:45:17.6618994Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6619170Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6619172Z 
2025-12-04T09:45:17.6619257Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6619521Z FAILED [3.8267s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp3ho9z2ol/flex_attention_configs.json was not created
2025-12-04T09:45:17.6619523Z 
2025-12-04T09:45:17.6619594Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6619748Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6619760Z 
2025-12-04T09:45:17.6619844Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6620118Z FAILED [4.0667s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp081kh257/flex_attention_configs.json was not created
2025-12-04T09:45:17.6620120Z 
2025-12-04T09:45:17.6620190Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6620348Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6620350Z 
2025-12-04T09:45:17.6620481Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6620745Z FAILED [4.1450s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpxkkpz8q7/flex_attention_configs.json was not created
2025-12-04T09:45:17.6620748Z 
2025-12-04T09:45:17.6620817Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6620976Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6620979Z 
2025-12-04T09:45:17.6621061Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6621332Z FAILED [3.9231s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp9dxvkrhw/flex_attention_configs.json was not created
2025-12-04T09:45:17.6621334Z 
2025-12-04T09:45:17.6621405Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6621561Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6621565Z 
2025-12-04T09:45:17.6621649Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6621912Z FAILED [3.9786s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpx5r9jo87/flex_attention_configs.json was not created
2025-12-04T09:45:17.6621916Z 
2025-12-04T09:45:17.6621987Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6622142Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6622143Z 
2025-12-04T09:45:17.6622228Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6622493Z FAILED [4.2159s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpsuzkclcu/flex_attention_configs.json was not created
2025-12-04T09:45:17.6622496Z 
2025-12-04T09:45:17.6622567Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6622722Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6622725Z 
2025-12-04T09:45:17.6622820Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6623084Z FAILED [4.3649s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpvzrmqh1r/flex_attention_configs.json was not created
2025-12-04T09:45:17.6623086Z 
2025-12-04T09:45:17.6623156Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6623313Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6623315Z 
2025-12-04T09:45:17.6623398Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6623678Z FAILED [4.2806s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp6to3xt_d/flex_attention_configs.json was not created
2025-12-04T09:45:17.6623692Z 
2025-12-04T09:45:17.6623763Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6623919Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6623921Z 
2025-12-04T09:45:17.6624005Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6624280Z FAILED [4.4002s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpt6axgtf_/flex_attention_configs.json was not created
2025-12-04T09:45:17.6624283Z 
2025-12-04T09:45:17.6624355Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6624509Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6624511Z 
2025-12-04T09:45:17.6624595Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6624859Z FAILED [3.7822s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpkqvme528/flex_attention_configs.json was not created
2025-12-04T09:45:17.6624861Z 
2025-12-04T09:45:17.6624931Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6625088Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6625090Z 
2025-12-04T09:45:17.6625172Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6625438Z FAILED [3.7276s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp7xa3g518/flex_attention_configs.json was not created
2025-12-04T09:45:17.6625441Z 
2025-12-04T09:45:17.6625511Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6625668Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6625670Z 
2025-12-04T09:45:17.6625754Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6626029Z FAILED [4.5690s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp5m0xxtnv/flex_attention_configs.json was not created
2025-12-04T09:45:17.6626031Z 
2025-12-04T09:45:17.6626100Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6626258Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6626261Z 
2025-12-04T09:45:17.6626344Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6626615Z FAILED [4.1873s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpukvz7181/flex_attention_configs.json was not created
2025-12-04T09:45:17.6626617Z 
2025-12-04T09:45:17.6626689Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6626844Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6626846Z 
2025-12-04T09:45:17.6626931Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6627195Z FAILED [4.0301s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpu5r6r1gp/flex_attention_configs.json was not created
2025-12-04T09:45:17.6627218Z 
2025-12-04T09:45:17.6627290Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6627446Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6627448Z 
2025-12-04T09:45:17.6627531Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6627807Z FAILED [3.8084s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpzrs9u7ki/flex_attention_configs.json was not created
2025-12-04T09:45:17.6627809Z 
2025-12-04T09:45:17.6627879Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6628036Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6628039Z 
2025-12-04T09:45:17.6628122Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6628389Z FAILED [4.0913s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpo_d0k7ct/flex_attention_configs.json was not created
2025-12-04T09:45:17.6628391Z 
2025-12-04T09:45:17.6628460Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6628615Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6628617Z 
2025-12-04T09:45:17.6628703Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6628965Z FAILED [4.1417s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpfx5o9jqp/flex_attention_configs.json was not created
2025-12-04T09:45:17.6628969Z 
2025-12-04T09:45:17.6629041Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6629198Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6629200Z 
2025-12-04T09:45:17.6629285Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6629549Z FAILED [4.1571s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpb10711_l/flex_attention_configs.json was not created
2025-12-04T09:45:17.6629552Z 
2025-12-04T09:45:17.6629623Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6629779Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6629782Z 
2025-12-04T09:45:17.6629864Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6630143Z FAILED [4.4631s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpfzimglfo/flex_attention_configs.json was not created
2025-12-04T09:45:17.6630145Z 
2025-12-04T09:45:17.6630215Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6630373Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6630374Z 
2025-12-04T09:45:17.6630493Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6630758Z FAILED [3.9384s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpqq3rq4tk/flex_attention_configs.json was not created
2025-12-04T09:45:17.6630775Z 
2025-12-04T09:45:17.6630845Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6631014Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6631016Z 
2025-12-04T09:45:17.6631101Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6631368Z FAILED [4.3235s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpa59m29km/flex_attention_configs.json was not created
2025-12-04T09:45:17.6631370Z 
2025-12-04T09:45:17.6631458Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6631613Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6631615Z 
2025-12-04T09:45:17.6631699Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6631961Z FAILED [3.9420s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp8_w3cgif/flex_attention_configs.json was not created
2025-12-04T09:45:17.6631963Z 
2025-12-04T09:45:17.6632035Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6632190Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6632194Z 
2025-12-04T09:45:17.6632276Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6632541Z FAILED [4.3835s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpv8xw9256/flex_attention_configs.json was not created
2025-12-04T09:45:17.6632544Z 
2025-12-04T09:45:17.6632615Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6632771Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6632775Z 
2025-12-04T09:45:17.6632859Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6633125Z FAILED [4.0163s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpcnpjpknz/flex_attention_configs.json was not created
2025-12-04T09:45:17.6633127Z 
2025-12-04T09:45:17.6633197Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6633353Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6633355Z 
2025-12-04T09:45:17.6633438Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6633699Z FAILED [4.0669s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp4_r1jh6s/flex_attention_configs.json was not created
2025-12-04T09:45:17.6633702Z 
2025-12-04T09:45:17.6633785Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6633940Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6633942Z 
2025-12-04T09:45:17.6634026Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6634291Z FAILED [4.3283s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpsimzr412/flex_attention_configs.json was not created
2025-12-04T09:45:17.6634293Z 
2025-12-04T09:45:17.6634364Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6634530Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6634543Z 
2025-12-04T09:45:17.6634626Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6634890Z FAILED [4.0865s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpun190rtr/flex_attention_configs.json was not created
2025-12-04T09:45:17.6634892Z 
2025-12-04T09:45:17.6634962Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6635130Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6635132Z 
2025-12-04T09:45:17.6635214Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6635479Z FAILED [4.5002s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpwzy0l12r/flex_attention_configs.json was not created
2025-12-04T09:45:17.6635482Z 
2025-12-04T09:45:17.6635551Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6635708Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6635710Z 
2025-12-04T09:45:17.6635794Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6636059Z FAILED [4.1988s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp0a8luhxf/flex_attention_configs.json was not created
2025-12-04T09:45:17.6636061Z 
2025-12-04T09:45:17.6636132Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6636288Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6636290Z 
2025-12-04T09:45:17.6636374Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6636636Z FAILED [4.3243s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp7_fehk8b/flex_attention_configs.json was not created
2025-12-04T09:45:17.6636638Z 
2025-12-04T09:45:17.6636709Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6636864Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6636867Z 
2025-12-04T09:45:17.6636949Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6637215Z FAILED [4.0704s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp2be0ko7i/flex_attention_configs.json was not created
2025-12-04T09:45:17.6637218Z 
2025-12-04T09:45:17.6637289Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6637457Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6637459Z 
2025-12-04T09:45:17.6637541Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6637806Z FAILED [4.6904s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpop9htqnm/flex_attention_configs.json was not created
2025-12-04T09:45:17.6637808Z 
2025-12-04T09:45:17.6637878Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6638035Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6638047Z 
2025-12-04T09:45:17.6638131Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6638405Z FAILED [4.1336s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpqm289lwi/flex_attention_configs.json was not created
2025-12-04T09:45:17.6638407Z 
2025-12-04T09:45:17.6638478Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6638633Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6638634Z 
2025-12-04T09:45:17.6638728Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6638992Z FAILED [3.9265s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp6v5mhi1a/flex_attention_configs.json was not created
2025-12-04T09:45:17.6638994Z 
2025-12-04T09:45:17.6639066Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6639222Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6639225Z 
2025-12-04T09:45:17.6639307Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6639578Z FAILED [4.3027s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpz0t24o3o/flex_attention_configs.json was not created
2025-12-04T09:45:17.6639580Z 
2025-12-04T09:45:17.6639651Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6639807Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6639810Z 
2025-12-04T09:45:17.6639894Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6640156Z FAILED [4.3690s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpr50o_zw3/flex_attention_configs.json was not created
2025-12-04T09:45:17.6640159Z 
2025-12-04T09:45:17.6640228Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6640384Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6640386Z 
2025-12-04T09:45:17.6640504Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6640766Z FAILED [4.2044s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmprsso7tvz/flex_attention_configs.json was not created
2025-12-04T09:45:17.6640769Z 
2025-12-04T09:45:17.6640840Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6640994Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6640997Z 
2025-12-04T09:45:17.6641093Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6641356Z FAILED [4.3931s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpo09mhc5r/flex_attention_configs.json was not created
2025-12-04T09:45:17.6641358Z 
2025-12-04T09:45:17.6641429Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6641585Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6641587Z 
2025-12-04T09:45:17.6641682Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6641954Z FAILED [4.3645s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmp35vgyqua/flex_attention_configs.json was not created
2025-12-04T09:45:17.6641970Z 
2025-12-04T09:45:17.6642041Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6642196Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6642198Z 
2025-12-04T09:45:17.6642280Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6642556Z FAILED [4.5234s] inductor/test_flex_attention.py::TestLearnableBiasesCUDA::test_flex_attention_logging_cuda - AssertionError: False is not true : Log file /tmp/tmpwj1h5tyv/flex_attention_configs.json was not created
2025-12-04T09:45:17.6642558Z 
2025-12-04T09:45:17.6642631Z To execute this test, run the following from the base repo dir:
2025-12-04T09:45:17.6642787Z     PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_flex_attention.py TestLearnableBiasesCUDA.test_flex_attention_logging_cuda
2025-12-04T09:45:17.6642790Z 
2025-12-04T09:45:17.6642874Z This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
2025-12-04T09:45:17.6642943Z =================== 49 failed, 1 passed in 208.71s (0:03:28) ===================
2025-12-04T09:45:17.6642945Z 
2025-12-04T09:45:17.6643121Z FINISHED PRINTING LOG FILE of inductor/test_flex_attention 1/4 (test/test-reports/inductor.test_flex_attention_1.4_1061c3085781a0ce_.log)
2025-12-04T09:45:17.6643124Z 
2025-12-04T09:45:17.6643243Z Finished inductor/test_flex_attention 1/4 ... [2025-12-04 09:45:15.395671][2246732.529608392], took 3.60min
2025-12-04T09:45:17.6643478Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:45:17.6643564Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:45:17.6643657Z GITHUB_RUN_ID, GITHUB_RUN_ATTEMPT, or ARTIFACTS_FILE_SUFFIX not set, not uploading
2025-12-04T09:45:17.6643708Z Uploading artifacts took 0.00 seconds
2025-12-04T09:45:17.6643759Z inductor/test_flex_attention 1/4 failed!
2025-12-04T09:45:17.6643866Z Running inductor/test_cutlass_backend 1/1 ... [2025-12-04 09:45:15.397074][2246732.531031618]
2025-12-04T09:45:17.6643915Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:45:17.6644291Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_cutlass_backend.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:45:15.397263]
2025-12-04T09:45:20.8508315Z 
2025-12-04T09:45:20.8508786Z inductor/test_cutlass_backend 1/1 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_cutlass_backend_1.1_5bc441ef42814b67_.log
2025-12-04T09:45:20.8509200Z Running 0 items in this shard:
2025-12-04T09:45:20.8509308Z 
2025-12-04T09:45:20.8509465Z Finished inductor/test_cutlass_backend 1/1 ... [2025-12-04 09:45:20.850506][2246737.984458102], took 0.09min
2025-12-04T09:45:20.8513420Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:45:20.8518816Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:45:20.8522284Z Running inductor/test_custom_op_autotune 1/1 ... [2025-12-04 09:45:20.852128][2246737.986085495]
2025-12-04T09:45:20.8522542Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:45:20.8524675Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_custom_op_autotune.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:45:20.852339]
2025-12-04T09:45:26.1961589Z 
2025-12-04T09:45:26.1962685Z inductor/test_custom_op_autotune 1/1 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_custom_op_autotune_1.1_03971ab8d73d5cda_.log
2025-12-04T09:45:26.1963306Z Running 0 items in this shard:
2025-12-04T09:45:26.1963451Z 
2025-12-04T09:45:26.1963703Z Finished inductor/test_custom_op_autotune 1/1 ... [2025-12-04 09:45:26.195811][2246743.329760484], took 0.09min
2025-12-04T09:45:26.1968837Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:45:26.1972669Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:45:26.1975104Z Running inductor/test_compile_subprocess 2/3 ... [2025-12-04 09:45:26.197417][2246743.331375587]
2025-12-04T09:45:26.1975435Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:45:26.1977053Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_compile_subprocess.py', '--shard-id=2', '--num-shards=3', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:45:26.197573]
2025-12-04T09:46:04.2372120Z 
2025-12-04T09:46:04.2373172Z inductor/test_compile_subprocess 2/3 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_compile_subprocess_2.3_f1aa0e75cae0db8b_.log
2025-12-04T09:46:04.2387331Z Running 50 items in this shard: test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda, test/inductor/test_compile_subprocess.py::GPUTests::test_dropout_deterministic_cuda
2025-12-04T09:46:04.2396342Z 
2025-12-04T09:46:04.2396540Z Finished inductor/test_compile_subprocess 2/3 ... [2025-12-04 09:46:04.238240][2246781.372193856], took 0.63min
2025-12-04T09:46:04.2397121Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:46:04.2397616Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:46:04.2397929Z Running dynamo/test_model_output 1/1 ... [2025-12-04 09:46:04.239562][2246781.373520204]
2025-12-04T09:46:04.2398185Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:46:04.2398827Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'dynamo/test_model_output.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:46:04.239721]
2025-12-04T09:46:06.6327375Z 
2025-12-04T09:46:06.6328157Z dynamo/test_model_output 1/1 was successful, full logs can be found in artifacts with path test/test-reports/dynamo.test_model_output_1.1_a0600735d94f03cb_.log
2025-12-04T09:46:06.6334826Z Running 50 items in this shard: test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr, test/dynamo/test_model_output.py::TestHFPretrained::test_pretrained_non_const_attr
2025-12-04T09:46:06.6340609Z 
2025-12-04T09:46:06.6340752Z Finished dynamo/test_model_output 1/1 ... [2025-12-04 09:46:06.632521][2246783.766474822], took 0.04min
2025-12-04T09:46:06.6341155Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:46:06.6341517Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:46:06.6341758Z Running inductor/test_selective_lowering 1/1 ... [2025-12-04 09:46:06.633908][2246783.767866379]
2025-12-04T09:46:06.6341961Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:46:06.6342438Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_selective_lowering.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:46:06.634094]
2025-12-04T09:46:12.4576425Z 
2025-12-04T09:46:12.4577314Z inductor/test_selective_lowering 1/1 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_selective_lowering_1.1_902229165a8abb3a_.log
2025-12-04T09:46:12.4577700Z Running 0 items in this shard:
2025-12-04T09:46:12.4577788Z 
2025-12-04T09:46:12.4577942Z Finished inductor/test_selective_lowering 1/1 ... [2025-12-04 09:46:12.457287][2246789.591241225], took 0.10min
2025-12-04T09:46:12.4582583Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:46:12.4584997Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:46:12.4587174Z Running dynamo/test_backends 1/1 ... [2025-12-04 09:46:12.458619][2246789.592575993]
2025-12-04T09:46:12.4587381Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:46:12.4589213Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'dynamo/test_backends.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:46:12.458798]
2025-12-04T09:46:18.4025702Z 
2025-12-04T09:46:18.4026642Z dynamo/test_backends 1/1 was successful, full logs can be found in artifacts with path test/test-reports/dynamo.test_backends_1.1_e2dca2c381d368bb_.log
2025-12-04T09:46:18.4027176Z Running 0 items in this shard:
2025-12-04T09:46:18.4027306Z 
2025-12-04T09:46:18.4027508Z Finished dynamo/test_backends 1/1 ... [2025-12-04 09:46:18.402304][2246795.536256394], took 0.10min
2025-12-04T09:46:18.4033369Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:46:18.4039429Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:46:18.4042428Z Running inductor/test_triton_heuristics 1/1 ... [2025-12-04 09:46:18.404117][2246795.538075254]
2025-12-04T09:46:18.4042736Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:46:18.4044589Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_triton_heuristics.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:46:18.404316]
2025-12-04T09:46:24.2415129Z 
2025-12-04T09:46:24.2416149Z inductor/test_triton_heuristics 1/1 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_triton_heuristics_1.1_eb681883ea525997_.log
2025-12-04T09:46:24.2416766Z Running 0 items in this shard:
2025-12-04T09:46:24.2416907Z 
2025-12-04T09:46:24.2417150Z Finished inductor/test_triton_heuristics 1/1 ... [2025-12-04 09:46:24.241214][2246801.375166976], took 0.10min
2025-12-04T09:46:24.2422350Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:46:24.2427546Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:46:24.2430667Z Running inductor/test_flex_decoding 2/2 ... [2025-12-04 09:46:24.242977][2246801.376933876]
2025-12-04T09:46:24.2430958Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:46:24.2433160Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_flex_decoding.py', '--shard-id=2', '--num-shards=2', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:46:24.243158]
2025-12-04T09:46:26.9113879Z 
2025-12-04T09:46:26.9114818Z inductor/test_flex_decoding 2/2 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_flex_decoding_2.2_b947e3330d826b28_.log
2025-12-04T09:46:26.9115974Z Running 0 items in this shard:
2025-12-04T09:46:26.9116092Z 
2025-12-04T09:46:26.9116308Z Finished inductor/test_flex_decoding 2/2 ... [2025-12-04 09:46:26.911116][2246804.045068655], took 0.04min
2025-12-04T09:46:26.9120565Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:46:26.9125895Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:46:26.9128822Z Running inductor/test_b2b_gemm 1/1 ... [2025-12-04 09:46:26.912798][2246804.046755197]
2025-12-04T09:46:26.9129081Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:46:26.9131629Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_b2b_gemm.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:46:26.913013]
2025-12-04T09:46:32.2648795Z 
2025-12-04T09:46:32.2649789Z inductor/test_b2b_gemm 1/1 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_b2b_gemm_1.1_f3a190b2a6506b09_.log
2025-12-04T09:46:32.2650227Z Running 0 items in this shard:
2025-12-04T09:46:32.2650344Z 
2025-12-04T09:46:32.2650597Z Finished inductor/test_b2b_gemm 1/1 ... [2025-12-04 09:46:32.264614][2246809.398565995], took 0.09min
2025-12-04T09:46:32.2655314Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:46:32.2661527Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:46:32.2664723Z Running export/test_unflatten 1/1 ... [2025-12-04 09:46:32.266387][2246809.400344675]
2025-12-04T09:46:32.2664978Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:46:32.2667214Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'export/test_unflatten.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:46:32.266606]
2025-12-04T09:46:34.1319736Z 
2025-12-04T09:46:34.1320626Z export/test_unflatten 1/1 was successful, full logs can be found in artifacts with path test/test-reports/export.test_unflatten_1.1_541be05c55d0546d_.log
2025-12-04T09:46:34.1320963Z Running 0 items in this shard:
2025-12-04T09:46:34.1321048Z 
2025-12-04T09:46:34.1321164Z Finished export/test_unflatten 1/1 ... [2025-12-04 09:46:34.131730][2246811.265682583], took 0.03min
2025-12-04T09:46:34.1324345Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:46:34.1328938Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:46:34.1331731Z Running export/test_hop 1/1 ... [2025-12-04 09:46:34.133086][2246811.26704412]
2025-12-04T09:46:34.1332355Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:46:34.1333720Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'export/test_hop.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:46:34.133266]
2025-12-04T09:46:36.9290199Z 
2025-12-04T09:46:36.9291079Z export/test_hop 1/1 was successful, full logs can be found in artifacts with path test/test-reports/export.test_hop_1.1_e2f428ff8d22b160_.log
2025-12-04T09:46:36.9291996Z Running 0 items in this shard:
2025-12-04T09:46:36.9292074Z 
2025-12-04T09:46:36.9292183Z Finished export/test_hop 1/1 ... [2025-12-04 09:46:36.928683][2246814.062636073], took 0.05min
2025-12-04T09:46:36.9296072Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:46:36.9303859Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:46:36.9305165Z Running export/test_serdes 1/1 ... [2025-12-04 09:46:36.930352][2246814.064310336]
2025-12-04T09:46:36.9305495Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:46:36.9307296Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'export/test_serdes.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:46:36.930553]
2025-12-04T09:46:43.3932026Z 
2025-12-04T09:46:43.3933000Z export/test_serdes 1/1 was successful, full logs can be found in artifacts with path test/test-reports/export.test_serdes_1.1_a2ede9bdf54b6543_.log
2025-12-04T09:46:43.3933568Z Running 0 items in this shard:
2025-12-04T09:46:43.3933714Z 
2025-12-04T09:46:43.3933924Z Finished export/test_serdes 1/1 ... [2025-12-04 09:46:43.392746][2246820.526698619], took 0.11min
2025-12-04T09:46:43.3938341Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:46:43.3943844Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:46:43.3946849Z Running inductor/test_debug_trace 1/1 ... [2025-12-04 09:46:43.394595][2246820.528551528]
2025-12-04T09:46:43.3947153Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:46:43.3949534Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'inductor/test_debug_trace.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:46:43.394822]
2025-12-04T09:46:49.0288950Z 
2025-12-04T09:46:49.0290228Z inductor/test_debug_trace 1/1 was successful, full logs can be found in artifacts with path test/test-reports/inductor.test_debug_trace_1.1_3fb9311045273a9a_.log
2025-12-04T09:46:49.0291176Z Running 0 items in this shard:
2025-12-04T09:46:49.0291388Z 
2025-12-04T09:46:49.0291706Z Finished inductor/test_debug_trace 1/1 ... [2025-12-04 09:46:49.028595][2246826.162547735], took 0.09min
2025-12-04T09:46:49.0296728Z Parsing testcases for test report: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/inductor.test_aot_inductor/inductor.test_aot_inductor-111eefd98bcfbfe3.xml
2025-12-04T09:46:49.0302376Z Failed to parse and upload json test reports: Unable to locate credentials
2025-12-04T09:46:49.0304749Z Running dynamo/test_guard_serialization 1/1 ... [2025-12-04 09:46:49.030382][2246826.164339375]
2025-12-04T09:46:49.0305107Z SCRIBE_GRAPHQL_ACCESS_TOKEN is NOT set
2025-12-04T09:46:49.0307676Z Executing ['/opt/conda/envs/py_3.12/bin/python', '-bb', 'dynamo/test_guard_serialization.py', '--shard-id=1', '--num-shards=1', '-v', '-vv', '-rfEX', '-p', 'no:xdist', '--use-pytest', '--flake-finder', '--flake-runs=50', '--import-slow-tests', '--import-disabled-tests', '--rerun-disabled-tests'] ... [2025-12-04 09:46:49.030594]
2025-12-04T13:53:01.7364871Z ##[error]The operation was canceled.
2025-12-04T13:53:01.7390898Z ##[group]Run # copy test results back to the mounted workspace, needed sudo, resulting permissions were correct
2025-12-04T13:53:01.7391210Z # copy test results back to the mounted workspace, needed sudo, resulting permissions were correct
2025-12-04T13:53:01.7391582Z docker exec -t "258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19" sh -c "cd ../pytorch && sudo cp -R test/test-reports ../workspace/test"
2025-12-04T13:53:01.7396040Z shell: /usr/bin/bash -e {0}
2025-12-04T13:53:01.7396150Z env:
2025-12-04T13:53:01.7396240Z   GIT_DEFAULT_BRANCH: main
2025-12-04T13:53:01.7396375Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T13:53:01.7396634Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T13:53:01.7396796Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T13:53:01.7397194Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T13:53:01.7397557Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T13:53:01.7397670Z   AWS_REGION: us-east-1
2025-12-04T13:53:01.7397850Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T13:53:01.7397998Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T13:53:01.7399907Z   AWS_SESSION_TOKEN: ***
2025-12-04T13:53:01.7400072Z   CONTAINER_NAME: 258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19
2025-12-04T13:53:01.7400251Z ##[endgroup]
2025-12-04T13:53:01.8007919Z ##[group]Run docker exec -t "258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19" sh -c "sudo chown -R 1001:1001 test"
2025-12-04T13:53:01.8008342Z docker exec -t "258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19" sh -c "sudo chown -R 1001:1001 test"
2025-12-04T13:53:01.8012897Z shell: /usr/bin/bash -e {0}
2025-12-04T13:53:01.8013015Z env:
2025-12-04T13:53:01.8013108Z   GIT_DEFAULT_BRANCH: main
2025-12-04T13:53:01.8013247Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T13:53:01.8013430Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T13:53:01.8013598Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T13:53:01.8013992Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T13:53:01.8014400Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T13:53:01.8014517Z   AWS_REGION: us-east-1
2025-12-04T13:53:01.8014708Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T13:53:01.8014862Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T13:53:01.8016836Z   AWS_SESSION_TOKEN: ***
2025-12-04T13:53:01.8017010Z   CONTAINER_NAME: 258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19
2025-12-04T13:53:01.8017194Z ##[endgroup]
2025-12-04T13:53:01.8773728Z ##[group]Run cat test/**/*_toprint.log || true
2025-12-04T13:53:01.8773891Z cat test/**/*_toprint.log || true
2025-12-04T13:53:01.8776806Z shell: /usr/bin/bash --noprofile --norc -e -o pipefail {0}
2025-12-04T13:53:01.8776957Z env:
2025-12-04T13:53:01.8777052Z   GIT_DEFAULT_BRANCH: main
2025-12-04T13:53:01.8777192Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T13:53:01.8777372Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T13:53:01.8777549Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T13:53:01.8777943Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T13:53:01.8778330Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T13:53:01.8778443Z   AWS_REGION: us-east-1
2025-12-04T13:53:01.8778601Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T13:53:01.8778763Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T13:53:01.8780710Z   AWS_SESSION_TOKEN: ***
2025-12-04T13:53:01.8780875Z   CONTAINER_NAME: 258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19
2025-12-04T13:53:01.8781051Z ##[endgroup]
2025-12-04T13:53:01.8834635Z Test results will be stored in test-reports/python-pytest/dynamo.test_guard_serialization/dynamo.test_guard_serialization-d61cec6039fe99bc.xml
2025-12-04T13:53:01.8834934Z ============================= test session starts ==============================
2025-12-04T13:53:01.8835228Z platform linux -- Python 3.12.5, pytest-7.3.2, pluggy-1.6.0 -- /opt/conda/envs/py_3.12/bin/python
2025-12-04T13:53:01.8835416Z cachedir: .pytest_cache
2025-12-04T13:53:01.8835688Z hypothesis profile 'pytorch_ci' -> database=None, max_examples=50, derandomize=True, suppress_health_check=[HealthCheck.too_slow]
2025-12-04T13:53:01.8836114Z rootdir: /var/lib/jenkins/pytorch
2025-12-04T13:53:01.8836242Z configfile: pytest.ini
2025-12-04T13:53:01.8836481Z plugins: hypothesis-6.56.4, cpp-2.3.0, flakefinder-1.1.0, rerunfailures-14.0, subtests-0.13.1, xdist-3.3.1, xdoctest-1.3.0, typeguard-4.3.0
2025-12-04T13:53:01.8836732Z collecting ... collected 56 items
2025-12-04T13:53:01.8836877Z stepcurrent: Cannot find last run test, not skipping
2025-12-04T13:53:01.8844668Z Running 50 items in this shard: test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module, test/dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module
2025-12-04T13:53:01.8852199Z 
2025-12-04T13:53:01.8852366Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.3305s] [  2%]
2025-12-04T13:53:01.8852718Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0891s] [  2%]
2025-12-04T13:53:01.8853110Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0820s] [  2%]
2025-12-04T13:53:01.8853458Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0834s] [  2%]
2025-12-04T13:53:01.8853803Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0708s] [  2%]
2025-12-04T13:53:01.8854148Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0774s] [  2%]
2025-12-04T13:53:01.8854492Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0798s] [  2%]
2025-12-04T13:53:01.8854840Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0872s] [  2%]
2025-12-04T13:53:01.8855185Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0862s] [  2%]
2025-12-04T13:53:01.8855534Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0792s] [  2%]
2025-12-04T13:53:01.8855878Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0989s] [  2%]
2025-12-04T13:53:01.8856223Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0761s] [  2%]
2025-12-04T13:53:01.8856567Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0873s] [  2%]
2025-12-04T13:53:01.8856927Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0870s] [  2%]
2025-12-04T13:53:01.8857288Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0789s] [  2%]
2025-12-04T13:53:01.8857636Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0963s] [  2%]
2025-12-04T13:53:01.8857985Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0751s] [  2%]
2025-12-04T13:53:01.8858329Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0885s] [  2%]
2025-12-04T13:53:01.8858676Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0862s] [  2%]
2025-12-04T13:53:01.8859020Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0802s] [  2%]
2025-12-04T13:53:01.8859367Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0939s] [  2%]
2025-12-04T13:53:01.8859712Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0900s] [  2%]
2025-12-04T13:53:01.8860063Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0936s] [  2%]
2025-12-04T13:53:01.8860458Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0412s] [  2%]
2025-12-04T13:53:01.8860805Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0842s] [  2%]
2025-12-04T13:53:01.8861153Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0860s] [  2%]
2025-12-04T13:53:01.8861502Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0991s] [  2%]
2025-12-04T13:53:01.8861849Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0839s] [  2%]
2025-12-04T13:53:01.8862200Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0866s] [  2%]
2025-12-04T13:53:01.8862583Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0835s] [  2%]
2025-12-04T13:53:01.8862929Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0838s] [  2%]
2025-12-04T13:53:01.8863279Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0871s] [  2%]
2025-12-04T13:53:01.8863628Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0797s] [  2%]
2025-12-04T13:53:01.8863976Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0706s] [  2%]
2025-12-04T13:53:01.8864324Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0758s] [  2%]
2025-12-04T13:53:01.8864673Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0758s] [  2%]
2025-12-04T13:53:01.8865019Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0995s] [  2%]
2025-12-04T13:53:01.8865367Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0866s] [  2%]
2025-12-04T13:53:01.8865712Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0860s] [  2%]
2025-12-04T13:53:01.8866060Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0778s] [  2%]
2025-12-04T13:53:01.8866450Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0861s] [  2%]
2025-12-04T13:53:01.8866811Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0891s] [  2%]
2025-12-04T13:53:01.8867165Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.1013s] [  2%]
2025-12-04T13:53:01.8867512Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0884s] [  2%]
2025-12-04T13:53:01.8867858Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0856s] [  2%]
2025-12-04T13:53:01.8868204Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0696s] [  2%]
2025-12-04T13:53:01.8868550Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0930s] [  2%]
2025-12-04T13:53:01.8868897Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0850s] [  2%]
2025-12-04T13:53:01.8869245Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0872s] [  2%]
2025-12-04T13:53:01.8869593Z dynamo/test_guard_serialization.py::TestGuardSerializationFSDP::test_guard_serialization_fsdp_module PASSED [0.0845s] [  2%]
2025-12-04T13:53:01.8869787Z 
2025-12-04T13:53:01.8870026Z - generated xml file: /var/lib/jenkins/pytorch/test/test-reports/python-pytest/dynamo.test_guard_serialization/dynamo.test_guard_serialization-d61cec6039fe99bc.xml -
2025-12-04T13:53:01.8870356Z ============================== 50 passed in 4.54s ==============================
2025-12-04T13:53:01.8927774Z Prepare all required actions
2025-12-04T13:53:01.8928087Z Getting action download info
2025-12-04T13:53:02.3207921Z Download action repository 'seemethere/upload-artifact-s3@v5' (SHA:baba72d0712b404f646cebe0730933554ebce96a)
2025-12-04T13:53:03.1662999Z Download action repository 'actions/upload-artifact@v4' (SHA:ea165f8d65b6e75b540449e92b4886f43607fa02)
2025-12-04T13:53:04.1104584Z ##[group]Run ./.github/actions/upload-test-artifacts
2025-12-04T13:53:04.1104733Z with:
2025-12-04T13:53:04.1104829Z   use-gha: true
2025-12-04T13:53:04.1104983Z   file-suffix: test-default-3-6-linux.rocm.gpu.gfx942.1.b_57116139325
2025-12-04T13:53:04.1105161Z   s3-bucket: gha-artifacts
2025-12-04T13:53:04.1105269Z env:
2025-12-04T13:53:04.1105360Z   GIT_DEFAULT_BRANCH: main
2025-12-04T13:53:04.1105498Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T13:53:04.1105674Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T13:53:04.1105866Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T13:53:04.1106250Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T13:53:04.1106639Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T13:53:04.1106755Z   AWS_REGION: us-east-1
2025-12-04T13:53:04.1106907Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T13:53:04.1107059Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T13:53:04.1108992Z   AWS_SESSION_TOKEN: ***
2025-12-04T13:53:04.1109163Z   CONTAINER_NAME: 258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19
2025-12-04T13:53:04.1109343Z ##[endgroup]
2025-12-04T13:53:04.1139468Z ##[group]Run actions/upload-artifact@v4
2025-12-04T13:53:04.1139601Z with:
2025-12-04T13:53:04.1139783Z   name: test-jsons-runattempt1-test-default-3-6-linux.rocm.gpu.gfx942.1.b_57116139325.zip
2025-12-04T13:53:04.1139991Z   retention-days: 14
2025-12-04T13:53:04.1140105Z   if-no-files-found: warn
2025-12-04T13:53:04.1140215Z   path: test/**/*.json
2025-12-04T13:53:04.1140407Z   compression-level: 6
2025-12-04T13:53:04.1140559Z   overwrite: false
2025-12-04T13:53:04.1140665Z   include-hidden-files: false
2025-12-04T13:53:04.1140783Z env:
2025-12-04T13:53:04.1140949Z   GIT_DEFAULT_BRANCH: main
2025-12-04T13:53:04.1141083Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T13:53:04.1141263Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T13:53:04.1141432Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T13:53:04.1141818Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T13:53:04.1142197Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T13:53:04.1142316Z   AWS_REGION: us-east-1
2025-12-04T13:53:04.1142447Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T13:53:04.1142601Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T13:53:04.1144523Z   AWS_SESSION_TOKEN: ***
2025-12-04T13:53:04.1144698Z   CONTAINER_NAME: 258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19
2025-12-04T13:53:04.1144878Z ##[endgroup]
2025-12-04T13:53:04.4818231Z With the provided path, there will be 6 files uploaded
2025-12-04T13:53:04.4821892Z Artifact name is valid!
2025-12-04T13:53:04.4823075Z Root directory input is valid!
2025-12-04T13:53:04.7044873Z Beginning upload of artifact content to blob storage
2025-12-04T13:53:05.0708881Z Uploaded bytes 46464
2025-12-04T13:53:05.1402341Z Finished uploading artifact content to blob storage!
2025-12-04T13:53:05.1403395Z SHA256 digest of uploaded artifact zip is b3c7d970c9d4af1e9afd806654a207c1c6b97323072602ffbec64541992b5d69
2025-12-04T13:53:05.1404029Z Finalizing artifact upload
2025-12-04T13:53:05.3462603Z Artifact test-jsons-runattempt1-test-default-3-6-linux.rocm.gpu.gfx942.1.b_57116139325.zip.zip successfully finalized. Artifact ID 4764759777
2025-12-04T13:53:05.3463707Z Artifact test-jsons-runattempt1-test-default-3-6-linux.rocm.gpu.gfx942.1.b_57116139325.zip has been successfully uploaded! Final size is 46464 bytes. Artifact ID is 4764759777
2025-12-04T13:53:05.3466906Z Artifact download URL: https://github.com/pytorch/pytorch/actions/runs/19922812470/artifacts/4764759777
2025-12-04T13:53:05.3590537Z ##[group]Run actions/upload-artifact@v4
2025-12-04T13:53:05.3590688Z with:
2025-12-04T13:53:05.3590876Z   name: test-reports-runattempt1-test-default-3-6-linux.rocm.gpu.gfx942.1.b_57116139325.zip
2025-12-04T13:53:05.3591086Z   retention-days: 14
2025-12-04T13:53:05.3591195Z   if-no-files-found: ignore
2025-12-04T13:53:05.3591314Z   path: test/**/*.xml
test/**/*.csv

2025-12-04T13:53:05.3591436Z   compression-level: 6
2025-12-04T13:53:05.3591539Z   overwrite: false
2025-12-04T13:53:05.3591643Z   include-hidden-files: false
2025-12-04T13:53:05.3591755Z env:
2025-12-04T13:53:05.3591847Z   GIT_DEFAULT_BRANCH: main
2025-12-04T13:53:05.3591989Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T13:53:05.3592168Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T13:53:05.3592341Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T13:53:05.3592728Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T13:53:05.3593099Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T13:53:05.3593217Z   AWS_REGION: us-east-1
2025-12-04T13:53:05.3593400Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T13:53:05.3593553Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T13:53:05.3595464Z   AWS_SESSION_TOKEN: ***
2025-12-04T13:53:05.3595634Z   CONTAINER_NAME: 258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19
2025-12-04T13:53:05.3595813Z ##[endgroup]
2025-12-04T13:53:05.7611507Z With the provided path, there will be 26 files uploaded
2025-12-04T13:53:05.7614434Z Artifact name is valid!
2025-12-04T13:53:05.7615031Z Root directory input is valid!
2025-12-04T13:53:05.9716293Z Beginning upload of artifact content to blob storage
2025-12-04T13:53:06.7066077Z Uploaded bytes 559442
2025-12-04T13:53:06.7847015Z Finished uploading artifact content to blob storage!
2025-12-04T13:53:06.7848287Z SHA256 digest of uploaded artifact zip is b3d38dd35ccb2729fa2d13ff63829ae648cde44c133b5bad4bafac37f8de5a69
2025-12-04T13:53:06.7849175Z Finalizing artifact upload
2025-12-04T13:53:07.0073019Z Artifact test-reports-runattempt1-test-default-3-6-linux.rocm.gpu.gfx942.1.b_57116139325.zip.zip successfully finalized. Artifact ID 4764760084
2025-12-04T13:53:07.0074427Z Artifact test-reports-runattempt1-test-default-3-6-linux.rocm.gpu.gfx942.1.b_57116139325.zip has been successfully uploaded! Final size is 559442 bytes. Artifact ID is 4764760084
2025-12-04T13:53:07.0079309Z Artifact download URL: https://github.com/pytorch/pytorch/actions/runs/19922812470/artifacts/4764760084
2025-12-04T13:53:07.0206477Z ##[group]Run actions/upload-artifact@v4
2025-12-04T13:53:07.0206659Z with:
2025-12-04T13:53:07.0206857Z   name: logs-runattempt1-test-default-3-6-linux.rocm.gpu.gfx942.1.b_57116139325.zip
2025-12-04T13:53:07.0207090Z   retention-days: 14
2025-12-04T13:53:07.0207224Z   if-no-files-found: ignore
2025-12-04T13:53:07.0207369Z   path: usage_log.txt
test/**/*.log

2025-12-04T13:53:07.0207510Z   compression-level: 6
2025-12-04T13:53:07.0207616Z   overwrite: false
2025-12-04T13:53:07.0207723Z   include-hidden-files: false
2025-12-04T13:53:07.0207839Z env:
2025-12-04T13:53:07.0207934Z   GIT_DEFAULT_BRANCH: main
2025-12-04T13:53:07.0208077Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T13:53:07.0208262Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T13:53:07.0208430Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T13:53:07.0208952Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T13:53:07.0209331Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T13:53:07.0209452Z   AWS_REGION: us-east-1
2025-12-04T13:53:07.0209640Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T13:53:07.0209798Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T13:53:07.0211989Z   AWS_SESSION_TOKEN: ***
2025-12-04T13:53:07.0212163Z   CONTAINER_NAME: 258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19
2025-12-04T13:53:07.0212345Z ##[endgroup]
2025-12-04T13:53:07.3923533Z Multiple search paths detected. Calculating the least common ancestor of all paths
2025-12-04T13:53:07.3924505Z The least common ancestor is /home/runner/_work/pytorch/pytorch. This will be the root directory of the artifact
2025-12-04T13:53:07.3924992Z With the provided path, there will be 24 files uploaded
2025-12-04T13:53:07.3927227Z Artifact name is valid!
2025-12-04T13:53:07.3927724Z Root directory input is valid!
2025-12-04T13:53:07.6135754Z Beginning upload of artifact content to blob storage
2025-12-04T13:53:08.2304485Z Uploaded bytes 538640
2025-12-04T13:53:08.2945466Z Finished uploading artifact content to blob storage!
2025-12-04T13:53:08.2946729Z SHA256 digest of uploaded artifact zip is f14937ef8466c36067dccf71b483749a80f5107d30c4464676c13febfc607dfd
2025-12-04T13:53:08.2947491Z Finalizing artifact upload
2025-12-04T13:53:08.7242806Z Artifact logs-runattempt1-test-default-3-6-linux.rocm.gpu.gfx942.1.b_57116139325.zip.zip successfully finalized. Artifact ID 4764760378
2025-12-04T13:53:08.7243404Z Artifact logs-runattempt1-test-default-3-6-linux.rocm.gpu.gfx942.1.b_57116139325.zip has been successfully uploaded! Final size is 538640 bytes. Artifact ID is 4764760378
2025-12-04T13:53:08.7249375Z Artifact download URL: https://github.com/pytorch/pytorch/actions/runs/19922812470/artifacts/4764760378
2025-12-04T13:53:08.7365393Z ##[group]Run # shellcheck disable=SC2156
2025-12-04T13:53:08.7365626Z # shellcheck disable=SC2156
2025-12-04T13:53:08.7365919Z find . -iname "core.[1-9]*" -exec docker exec "${CONTAINER_NAME}" sh -c "gdb python {} -ex 'bt' -ex 'q'" \;
2025-12-04T13:53:08.7370517Z shell: /usr/bin/bash -e {0}
2025-12-04T13:53:08.7370640Z env:
2025-12-04T13:53:08.7370742Z   GIT_DEFAULT_BRANCH: main
2025-12-04T13:53:08.7370969Z   RUNNER_ARTIFACT_DIR: /home/runner/_work/_temp/artifacts
2025-12-04T13:53:08.7371164Z   RUNNER_TEST_RESULTS_DIR: /home/runner/_work/_temp/test-results
2025-12-04T13:53:08.7371345Z   RUNNER_DOCS_DIR: /home/runner/_work/_temp/docs
2025-12-04T13:53:08.7371749Z   GPU_FLAG: --device=/dev/mem --device=/dev/kfd --group-add 110 --device /dev/dri/renderD152 --group-add video --group-add 109 --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host
2025-12-04T13:53:08.7372136Z   AWS_DEFAULT_REGION: us-east-1
2025-12-04T13:53:08.7372282Z   AWS_REGION: us-east-1
2025-12-04T13:53:08.7372511Z   AWS_ACCESS_KEY_ID: ***
2025-12-04T13:53:08.7372675Z   AWS_SECRET_ACCESS_KEY: ***
2025-12-04T13:53:08.7374675Z   AWS_SESSION_TOKEN: ***
2025-12-04T13:53:08.7374855Z   CONTAINER_NAME: 258f1619ff55f8d5b8f36e9d2cba17966f7fb641c5815f574be9334dbfc14a19
2025-12-04T13:53:08.7375052Z ##[endgroup]
2025-12-04T13:53:08.8713448Z Post job cleanup.
2025-12-04T13:53:08.8725564Z Post job cleanup.
2025-12-04T13:53:08.8919772Z Logging out of registry 308535385114.dkr.ecr.us-east-1.amazonaws.com
2025-12-04T13:53:08.9119393Z Post job cleanup.
2025-12-04T13:53:08.9737193Z Post job cleanup.
2025-12-04T13:53:08.9756473Z Post job cleanup.
2025-12-04T13:53:09.0202868Z [command]/usr/bin/git version
2025-12-04T13:53:09.0227356Z git version 2.52.0
2025-12-04T13:53:09.0247372Z Copying '/home/runner/.gitconfig' to '/home/runner/_work/_temp/e65673f7-14dc-4782-b53d-163efa622df2/.gitconfig'
2025-12-04T13:53:09.0253317Z Temporarily overriding HOME='/home/runner/_work/_temp/e65673f7-14dc-4782-b53d-163efa622df2' before making global git config changes
2025-12-04T13:53:09.0253815Z Adding repository directory to the temporary git global config as a safe directory
2025-12-04T13:53:09.0256221Z [command]/usr/bin/git config --global --add safe.directory /home/runner/_work/pytorch/pytorch
2025-12-04T13:53:09.0290297Z [command]/usr/bin/git config --local --name-only --get-regexp core\.sshCommand
2025-12-04T13:53:09.0316636Z [command]/usr/bin/git submodule foreach --recursive sh -c "git config --local --name-only --get-regexp 'core\.sshCommand' && git config --local --unset-all 'core.sshCommand' || :"
2025-12-04T13:53:09.0513402Z Entering 'android/libs/fbjni'
2025-12-04T13:53:09.0559481Z Entering 'third_party/FP16'
2025-12-04T13:53:09.0592578Z Entering 'third_party/FXdiv'
2025-12-04T13:53:09.0631070Z Entering 'third_party/NNPACK'
2025-12-04T13:53:09.0669086Z Entering 'third_party/NVTX'
2025-12-04T13:53:09.0697213Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T13:53:09.0728877Z Entering 'third_party/XNNPACK'
2025-12-04T13:53:09.0766254Z Entering 'third_party/aiter'
2025-12-04T13:53:09.0796610Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T13:53:09.0825111Z Entering 'third_party/benchmark'
2025-12-04T13:53:09.0852271Z Entering 'third_party/composable_kernel'
2025-12-04T13:53:09.0884156Z Entering 'third_party/cpp-httplib'
2025-12-04T13:53:09.0911151Z Entering 'third_party/cpuinfo'
2025-12-04T13:53:09.0935912Z Entering 'third_party/cudnn_frontend'
2025-12-04T13:53:09.0964838Z Entering 'third_party/cutlass'
2025-12-04T13:53:09.0996808Z Entering 'third_party/fbgemm'
2025-12-04T13:53:09.1022741Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T13:53:09.1044912Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T13:53:09.1079776Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T13:53:09.1105395Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T13:53:09.1134582Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T13:53:09.1160933Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T13:53:09.1182965Z Entering 'third_party/fbgemm/external/json'
2025-12-04T13:53:09.1205566Z Entering 'third_party/flash-attention'
2025-12-04T13:53:09.1229639Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T13:53:09.1258533Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T13:53:09.1287196Z Entering 'third_party/flatbuffers'
2025-12-04T13:53:09.1318287Z Entering 'third_party/fmt'
2025-12-04T13:53:09.1342573Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T13:53:09.1367121Z Entering 'third_party/gloo'
2025-12-04T13:53:09.1396658Z Entering 'third_party/googletest'
2025-12-04T13:53:09.1421724Z Entering 'third_party/ideep'
2025-12-04T13:53:09.1445918Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T13:53:09.1478019Z Entering 'third_party/ittapi'
2025-12-04T13:53:09.1502338Z Entering 'third_party/kineto'
2025-12-04T13:53:09.1527303Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T13:53:09.1557697Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T13:53:09.1580459Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T13:53:09.1602216Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T13:53:09.1629298Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T13:53:09.1660804Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T13:53:09.1691019Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T13:53:09.1715296Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T13:53:09.1738972Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T13:53:09.1768793Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T13:53:09.1795428Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T13:53:09.1822322Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T13:53:09.1854521Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T13:53:09.1882705Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T13:53:09.1902202Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T13:53:09.1926810Z Entering 'third_party/kleidiai'
2025-12-04T13:53:09.1956144Z Entering 'third_party/mimalloc'
2025-12-04T13:53:09.1980620Z Entering 'third_party/nlohmann'
2025-12-04T13:53:09.2006399Z Entering 'third_party/onnx'
2025-12-04T13:53:09.2042843Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T13:53:09.2071231Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T13:53:09.2096411Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T13:53:09.2126860Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T13:53:09.2150725Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T13:53:09.2175113Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T13:53:09.2200888Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T13:53:09.2222490Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T13:53:09.2243420Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T13:53:09.2265361Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T13:53:09.2288767Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T13:53:09.2315744Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T13:53:09.2355420Z Entering 'third_party/pocketfft'
2025-12-04T13:53:09.2383736Z Entering 'third_party/protobuf'
2025-12-04T13:53:09.2410137Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T13:53:09.2435231Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T13:53:09.2460962Z Entering 'third_party/psimd'
2025-12-04T13:53:09.2486065Z Entering 'third_party/pthreadpool'
2025-12-04T13:53:09.2510558Z Entering 'third_party/pybind11'
2025-12-04T13:53:09.2533671Z Entering 'third_party/python-peachpy'
2025-12-04T13:53:09.2557919Z Entering 'third_party/sleef'
2025-12-04T13:53:09.2583554Z Entering 'third_party/tensorpipe'
2025-12-04T13:53:09.2609169Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T13:53:09.2634891Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T13:53:09.2657203Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T13:53:09.2682825Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T13:53:09.2703491Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T13:53:09.2751535Z [command]/usr/bin/git config --local --name-only --get-regexp http\.https\:\/\/github\.com\/\.extraheader
2025-12-04T13:53:09.2770509Z http.https://github.com/.extraheader
2025-12-04T13:53:09.2784874Z [command]/usr/bin/git config --local --unset-all http.https://github.com/.extraheader
2025-12-04T13:53:09.2804685Z [command]/usr/bin/git submodule foreach --recursive sh -c "git config --local --name-only --get-regexp 'http\.https\:\/\/github\.com\/\.extraheader' && git config --local --unset-all 'http.https://github.com/.extraheader' || :"
2025-12-04T13:53:09.3012735Z Entering 'android/libs/fbjni'
2025-12-04T13:53:09.3029975Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3049507Z Entering 'third_party/FP16'
2025-12-04T13:53:09.3068116Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3094319Z Entering 'third_party/FXdiv'
2025-12-04T13:53:09.3111832Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3131570Z Entering 'third_party/NNPACK'
2025-12-04T13:53:09.3149435Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3169497Z Entering 'third_party/NVTX'
2025-12-04T13:53:09.3188632Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3207546Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T13:53:09.3225386Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3247553Z Entering 'third_party/XNNPACK'
2025-12-04T13:53:09.3267228Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3293852Z Entering 'third_party/aiter'
2025-12-04T13:53:09.3311909Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3333392Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T13:53:09.3349347Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3375162Z Entering 'third_party/benchmark'
2025-12-04T13:53:09.3395333Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3415470Z Entering 'third_party/composable_kernel'
2025-12-04T13:53:09.3435902Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3463214Z Entering 'third_party/cpp-httplib'
2025-12-04T13:53:09.3481380Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3500483Z Entering 'third_party/cpuinfo'
2025-12-04T13:53:09.3516334Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3538162Z Entering 'third_party/cudnn_frontend'
2025-12-04T13:53:09.3562209Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3582055Z Entering 'third_party/cutlass'
2025-12-04T13:53:09.3597067Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3619340Z Entering 'third_party/fbgemm'
2025-12-04T13:53:09.3643588Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3663467Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T13:53:09.3680007Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3706477Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T13:53:09.3726262Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3750569Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T13:53:09.3766644Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3792224Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T13:53:09.3809297Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3836056Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T13:53:09.3850787Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3869452Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T13:53:09.3888171Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3909554Z Entering 'third_party/fbgemm/external/json'
2025-12-04T13:53:09.3927396Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3951183Z Entering 'third_party/flash-attention'
2025-12-04T13:53:09.3971867Z http.https://github.com/.extraheader
2025-12-04T13:53:09.3996028Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T13:53:09.4010130Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4039814Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T13:53:09.4056897Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4088380Z Entering 'third_party/flatbuffers'
2025-12-04T13:53:09.4103007Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4126084Z Entering 'third_party/fmt'
2025-12-04T13:53:09.4143752Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4163461Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T13:53:09.4178507Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4201661Z Entering 'third_party/gloo'
2025-12-04T13:53:09.4215765Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4236490Z Entering 'third_party/googletest'
2025-12-04T13:53:09.4257427Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4278675Z Entering 'third_party/ideep'
2025-12-04T13:53:09.4297845Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4316195Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T13:53:09.4333614Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4361329Z Entering 'third_party/ittapi'
2025-12-04T13:53:09.4376615Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4398070Z Entering 'third_party/kineto'
2025-12-04T13:53:09.4418193Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4440694Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T13:53:09.4460763Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4479932Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T13:53:09.4497716Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4521179Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T13:53:09.4539430Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4558742Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T13:53:09.4574092Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4593153Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T13:53:09.4610890Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4630092Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T13:53:09.4649373Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4672581Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T13:53:09.4687429Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4704705Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T13:53:09.4719457Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4736755Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T13:53:09.4755228Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4776104Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T13:53:09.4793991Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4820238Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T13:53:09.4836347Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4861381Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T13:53:09.4881885Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4899047Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T13:53:09.4915669Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4941036Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T13:53:09.4962608Z http.https://github.com/.extraheader
2025-12-04T13:53:09.4987702Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T13:53:09.5007918Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5033748Z Entering 'third_party/kleidiai'
2025-12-04T13:53:09.5060255Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5081047Z Entering 'third_party/mimalloc'
2025-12-04T13:53:09.5098357Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5122200Z Entering 'third_party/nlohmann'
2025-12-04T13:53:09.5141107Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5166405Z Entering 'third_party/onnx'
2025-12-04T13:53:09.5189794Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5221915Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T13:53:09.5242927Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5265376Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T13:53:09.5280823Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5301617Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T13:53:09.5317167Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5339539Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T13:53:09.5364750Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5387653Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T13:53:09.5404946Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5429074Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T13:53:09.5452674Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5472701Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T13:53:09.5490537Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5512312Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T13:53:09.5528114Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5551092Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T13:53:09.5571959Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5600375Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T13:53:09.5617060Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5636718Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T13:53:09.5653426Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5682014Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T13:53:09.5697698Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5726870Z Entering 'third_party/pocketfft'
2025-12-04T13:53:09.5745638Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5764414Z Entering 'third_party/protobuf'
2025-12-04T13:53:09.5780460Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5803171Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T13:53:09.5825065Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5846879Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T13:53:09.5861469Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5887847Z Entering 'third_party/psimd'
2025-12-04T13:53:09.5903077Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5925441Z Entering 'third_party/pthreadpool'
2025-12-04T13:53:09.5942783Z http.https://github.com/.extraheader
2025-12-04T13:53:09.5962626Z Entering 'third_party/pybind11'
2025-12-04T13:53:09.5982362Z http.https://github.com/.extraheader
2025-12-04T13:53:09.6003792Z Entering 'third_party/python-peachpy'
2025-12-04T13:53:09.6024251Z http.https://github.com/.extraheader
2025-12-04T13:53:09.6044221Z Entering 'third_party/sleef'
2025-12-04T13:53:09.6060120Z http.https://github.com/.extraheader
2025-12-04T13:53:09.6084764Z Entering 'third_party/tensorpipe'
2025-12-04T13:53:09.6102250Z http.https://github.com/.extraheader
2025-12-04T13:53:09.6121607Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T13:53:09.6144417Z http.https://github.com/.extraheader
2025-12-04T13:53:09.6161792Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T13:53:09.6182528Z http.https://github.com/.extraheader
2025-12-04T13:53:09.6203739Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T13:53:09.6221328Z http.https://github.com/.extraheader
2025-12-04T13:53:09.6243176Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T13:53:09.6259777Z http.https://github.com/.extraheader
2025-12-04T13:53:09.6278099Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T13:53:09.6294813Z http.https://github.com/.extraheader
2025-12-04T13:53:09.6348144Z [command]/usr/bin/git config --local --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.6370395Z [command]/usr/bin/git submodule foreach --recursive git config --local --show-origin --name-only --get-regexp remote.origin.url
2025-12-04T13:53:09.6550009Z Entering 'android/libs/fbjni'
2025-12-04T13:53:09.6562385Z file:/home/runner/_work/pytorch/pytorch/.git/modules/android/libs/fbjni/config	remote.origin.url
2025-12-04T13:53:09.6576137Z Entering 'third_party/FP16'
2025-12-04T13:53:09.6589100Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FP16/config	remote.origin.url
2025-12-04T13:53:09.6604010Z Entering 'third_party/FXdiv'
2025-12-04T13:53:09.6613764Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FXdiv/config	remote.origin.url
2025-12-04T13:53:09.6625333Z Entering 'third_party/NNPACK'
2025-12-04T13:53:09.6639016Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK/config	remote.origin.url
2025-12-04T13:53:09.6647700Z Entering 'third_party/NVTX'
2025-12-04T13:53:09.6658605Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NVTX/config	remote.origin.url
2025-12-04T13:53:09.6667221Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T13:53:09.6677768Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/VulkanMemoryAllocator/config	remote.origin.url
2025-12-04T13:53:09.6686251Z Entering 'third_party/XNNPACK'
2025-12-04T13:53:09.6699819Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/XNNPACK/config	remote.origin.url
2025-12-04T13:53:09.6714810Z Entering 'third_party/aiter'
2025-12-04T13:53:09.6724442Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/config	remote.origin.url
2025-12-04T13:53:09.6736750Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T13:53:09.6753891Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/modules/3rdparty/composable_kernel/config	remote.origin.url
2025-12-04T13:53:09.6767845Z Entering 'third_party/benchmark'
2025-12-04T13:53:09.6784348Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T13:53:09.6793081Z Entering 'third_party/composable_kernel'
2025-12-04T13:53:09.6803329Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/composable_kernel/config	remote.origin.url
2025-12-04T13:53:09.6815180Z Entering 'third_party/cpp-httplib'
2025-12-04T13:53:09.6825896Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpp-httplib/config	remote.origin.url
2025-12-04T13:53:09.6840233Z Entering 'third_party/cpuinfo'
2025-12-04T13:53:09.6850759Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpuinfo/config	remote.origin.url
2025-12-04T13:53:09.6860968Z Entering 'third_party/cudnn_frontend'
2025-12-04T13:53:09.6871119Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cudnn_frontend/config	remote.origin.url
2025-12-04T13:53:09.6882553Z Entering 'third_party/cutlass'
2025-12-04T13:53:09.6895738Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cutlass/config	remote.origin.url
2025-12-04T13:53:09.6913258Z Entering 'third_party/fbgemm'
2025-12-04T13:53:09.6928374Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/config	remote.origin.url
2025-12-04T13:53:09.6939643Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T13:53:09.6952045Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/asmjit/config	remote.origin.url
2025-12-04T13:53:09.6961767Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T13:53:09.6975499Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/composable_kernel/config	remote.origin.url
2025-12-04T13:53:09.6988281Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T13:53:09.7002562Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cpuinfo/config	remote.origin.url
2025-12-04T13:53:09.7016553Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T13:53:09.7027907Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cutlass/config	remote.origin.url
2025-12-04T13:53:09.7044249Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T13:53:09.7054766Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/googletest/config	remote.origin.url
2025-12-04T13:53:09.7065670Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T13:53:09.7077705Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/hipify_torch/config	remote.origin.url
2025-12-04T13:53:09.7088867Z Entering 'third_party/fbgemm/external/json'
2025-12-04T13:53:09.7100882Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/json/config	remote.origin.url
2025-12-04T13:53:09.7113296Z Entering 'third_party/flash-attention'
2025-12-04T13:53:09.7124985Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/config	remote.origin.url
2025-12-04T13:53:09.7137989Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T13:53:09.7149146Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/composable_kernel/config	remote.origin.url
2025-12-04T13:53:09.7162352Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T13:53:09.7173322Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/cutlass/config	remote.origin.url
2025-12-04T13:53:09.7187261Z Entering 'third_party/flatbuffers'
2025-12-04T13:53:09.7200116Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flatbuffers/config	remote.origin.url
2025-12-04T13:53:09.7210176Z Entering 'third_party/fmt'
2025-12-04T13:53:09.7228797Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fmt/config	remote.origin.url
2025-12-04T13:53:09.7241860Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T13:53:09.7256587Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/gemmlowp/gemmlowp/config	remote.origin.url
2025-12-04T13:53:09.7270008Z Entering 'third_party/gloo'
2025-12-04T13:53:09.7283288Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/gloo/config	remote.origin.url
2025-12-04T13:53:09.7295547Z Entering 'third_party/googletest'
2025-12-04T13:53:09.7305709Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/googletest/config	remote.origin.url
2025-12-04T13:53:09.7317817Z Entering 'third_party/ideep'
2025-12-04T13:53:09.7329040Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/config	remote.origin.url
2025-12-04T13:53:09.7339585Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T13:53:09.7351304Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/modules/mkl-dnn/config	remote.origin.url
2025-12-04T13:53:09.7364833Z Entering 'third_party/ittapi'
2025-12-04T13:53:09.7374604Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ittapi/config	remote.origin.url
2025-12-04T13:53:09.7384214Z Entering 'third_party/kineto'
2025-12-04T13:53:09.7394632Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/config	remote.origin.url
2025-12-04T13:53:09.7404469Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T13:53:09.7416866Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/config	remote.origin.url
2025-12-04T13:53:09.7426788Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T13:53:09.7436532Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/DCGM/config	remote.origin.url
2025-12-04T13:53:09.7445470Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T13:53:09.7455831Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/cpr/config	remote.origin.url
2025-12-04T13:53:09.7463781Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T13:53:09.7475719Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/fmt/config	remote.origin.url
2025-12-04T13:53:09.7483824Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T13:53:09.7493306Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/config	remote.origin.url
2025-12-04T13:53:09.7500836Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T13:53:09.7514678Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/modules/doc/config	remote.origin.url
2025-12-04T13:53:09.7531610Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T13:53:09.7543296Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/glog/config	remote.origin.url
2025-12-04T13:53:09.7552082Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T13:53:09.7567500Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/googletest/config	remote.origin.url
2025-12-04T13:53:09.7576255Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T13:53:09.7588014Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/json/config	remote.origin.url
2025-12-04T13:53:09.7597118Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T13:53:09.7606703Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/pfs/config	remote.origin.url
2025-12-04T13:53:09.7614848Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T13:53:09.7623807Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/config	remote.origin.url
2025-12-04T13:53:09.7631751Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T13:53:09.7647578Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/civetweb/config	remote.origin.url
2025-12-04T13:53:09.7659203Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T13:53:09.7669313Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/googletest/config	remote.origin.url
2025-12-04T13:53:09.7681528Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T13:53:09.7692060Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/fmt/config	remote.origin.url
2025-12-04T13:53:09.7700828Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T13:53:09.7712383Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/googletest/config	remote.origin.url
2025-12-04T13:53:09.7724223Z Entering 'third_party/kleidiai'
2025-12-04T13:53:09.7736051Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kleidiai/config	remote.origin.url
2025-12-04T13:53:09.7745658Z Entering 'third_party/mimalloc'
2025-12-04T13:53:09.7759143Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/mimalloc/config	remote.origin.url
2025-12-04T13:53:09.7779587Z Entering 'third_party/nlohmann'
2025-12-04T13:53:09.7781375Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/nlohmann/config	remote.origin.url
2025-12-04T13:53:09.7791293Z Entering 'third_party/onnx'
2025-12-04T13:53:09.7801192Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/config	remote.origin.url
2025-12-04T13:53:09.7816726Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T13:53:09.7829875Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T13:53:09.7841883Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T13:53:09.7854254Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/config	remote.origin.url
2025-12-04T13:53:09.7868159Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T13:53:09.7881698Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T13:53:09.7890275Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T13:53:09.7908934Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/googletest/config	remote.origin.url
2025-12-04T13:53:09.7920483Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T13:53:09.7940259Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/ms-gsl/config	remote.origin.url
2025-12-04T13:53:09.7951134Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T13:53:09.7963033Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/nlohmann-json/config	remote.origin.url
2025-12-04T13:53:09.7971579Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T13:53:09.7985735Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentelemetry-proto/config	remote.origin.url
2025-12-04T13:53:09.7993742Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T13:53:09.8004183Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentracing-cpp/config	remote.origin.url
2025-12-04T13:53:09.8012733Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T13:53:09.8024578Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/config	remote.origin.url
2025-12-04T13:53:09.8033742Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T13:53:09.8049850Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/civetweb/config	remote.origin.url
2025-12-04T13:53:09.8061762Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T13:53:09.8072406Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/googletest/config	remote.origin.url
2025-12-04T13:53:09.8083665Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T13:53:09.8093514Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/tools/vcpkg/config	remote.origin.url
2025-12-04T13:53:09.8119837Z Entering 'third_party/pocketfft'
2025-12-04T13:53:09.8129918Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/pocketfft/config	remote.origin.url
2025-12-04T13:53:09.8140580Z Entering 'third_party/protobuf'
2025-12-04T13:53:09.8153984Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/config	remote.origin.url
2025-12-04T13:53:09.8165281Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T13:53:09.8183652Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T13:53:09.8193153Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T13:53:09.8211974Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/googletest/config	remote.origin.url
2025-12-04T13:53:09.8229898Z Entering 'third_party/psimd'
2025-12-04T13:53:09.8240819Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/psimd/config	remote.origin.url
2025-12-04T13:53:09.8252591Z Entering 'third_party/pthreadpool'
2025-12-04T13:53:09.8262800Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/pthreadpool/config	remote.origin.url
2025-12-04T13:53:09.8271384Z Entering 'third_party/pybind11'
2025-12-04T13:53:09.8280822Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T13:53:09.8289575Z Entering 'third_party/python-peachpy'
2025-12-04T13:53:09.8300493Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/python-peachpy/config	remote.origin.url
2025-12-04T13:53:09.8308998Z Entering 'third_party/sleef'
2025-12-04T13:53:09.8318316Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/sleef/config	remote.origin.url
2025-12-04T13:53:09.8330166Z Entering 'third_party/tensorpipe'
2025-12-04T13:53:09.8344853Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/config	remote.origin.url
2025-12-04T13:53:09.8356740Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T13:53:09.8368264Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/googletest/config	remote.origin.url
2025-12-04T13:53:09.8377945Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T13:53:09.8398546Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libnop/config	remote.origin.url
2025-12-04T13:53:09.8410301Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T13:53:09.8424403Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libuv/config	remote.origin.url
2025-12-04T13:53:09.8434737Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T13:53:09.8448384Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T13:53:09.8458865Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T13:53:09.8471315Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/modules/tools/clang/config	remote.origin.url
2025-12-04T13:53:09.8509189Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/android/libs/fbjni/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8527600Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FP16/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8546450Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FXdiv/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8562452Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8581029Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NVTX/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8595589Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/VulkanMemoryAllocator/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8611273Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/XNNPACK/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8625505Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8642126Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/modules/3rdparty/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8661319Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/benchmark/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8682969Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8699993Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpp-httplib/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8720364Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpuinfo/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8734888Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cudnn_frontend/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8751648Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cutlass/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8766201Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8780821Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/asmjit/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8795711Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8816624Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cpuinfo/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8834121Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cutlass/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8848437Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8864569Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/hipify_torch/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8878613Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/json/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8893233Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8907233Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8926379Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/cutlass/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8940848Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flatbuffers/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8956562Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fmt/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8972237Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/gemmlowp/gemmlowp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.8986184Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/gloo/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9001069Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9015082Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9029349Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/modules/mkl-dnn/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9043442Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/ittapi/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9058032Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9072669Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9087823Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/DCGM/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9103749Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/cpr/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9118495Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/fmt/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9134233Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9156202Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/modules/doc/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9172938Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/glog/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9192200Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9206730Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/json/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9221290Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/pfs/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9235674Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9249882Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/civetweb/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9264544Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9277880Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/fmt/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9296050Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9311401Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kleidiai/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9325382Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/mimalloc/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9338419Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/nlohmann/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9355077Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9369580Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/modules/third_party/pybind11/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9384949Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9399767Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/benchmark/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9414910Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9431524Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/ms-gsl/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9447227Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/nlohmann-json/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9462872Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentelemetry-proto/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9477956Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentracing-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9493248Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9508165Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/civetweb/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9523102Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9536600Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/tools/vcpkg/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9550971Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/pocketfft/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9565620Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9581166Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/benchmark/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9601603Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9617169Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/psimd/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9632606Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/pthreadpool/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9651571Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/pybind11/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9666054Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/python-peachpy/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9693279Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/sleef/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9709667Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9724473Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9739803Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libnop/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9754513Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libuv/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9769459Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9783809Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/modules/tools/clang/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:09.9880023Z Post job cleanup.
2025-12-04T13:53:10.0324388Z [command]/usr/bin/git version
2025-12-04T13:53:10.0353709Z git version 2.52.0
2025-12-04T13:53:10.0381694Z Copying '/home/runner/.gitconfig' to '/home/runner/_work/_temp/8e3fe175-4406-478a-b65c-dddda81e1fff/.gitconfig'
2025-12-04T13:53:10.0387631Z Temporarily overriding HOME='/home/runner/_work/_temp/8e3fe175-4406-478a-b65c-dddda81e1fff' before making global git config changes
2025-12-04T13:53:10.0388109Z Adding repository directory to the temporary git global config as a safe directory
2025-12-04T13:53:10.0390274Z [command]/usr/bin/git config --global --add safe.directory /home/runner/_work/pytorch/pytorch
2025-12-04T13:53:10.0410516Z [command]/usr/bin/git config --local --name-only --get-regexp core\.sshCommand
2025-12-04T13:53:10.0434742Z [command]/usr/bin/git submodule foreach --recursive sh -c "git config --local --name-only --get-regexp 'core\.sshCommand' && git config --local --unset-all 'core.sshCommand' || :"
2025-12-04T13:53:10.0605861Z Entering 'android/libs/fbjni'
2025-12-04T13:53:10.0652216Z Entering 'third_party/FP16'
2025-12-04T13:53:10.0679297Z Entering 'third_party/FXdiv'
2025-12-04T13:53:10.0704233Z Entering 'third_party/NNPACK'
2025-12-04T13:53:10.0739640Z Entering 'third_party/NVTX'
2025-12-04T13:53:10.0763438Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T13:53:10.0786892Z Entering 'third_party/XNNPACK'
2025-12-04T13:53:10.0815210Z Entering 'third_party/aiter'
2025-12-04T13:53:10.0837884Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T13:53:10.0865045Z Entering 'third_party/benchmark'
2025-12-04T13:53:10.0903310Z Entering 'third_party/composable_kernel'
2025-12-04T13:53:10.0948340Z Entering 'third_party/cpp-httplib'
2025-12-04T13:53:10.0980223Z Entering 'third_party/cpuinfo'
2025-12-04T13:53:10.1005613Z Entering 'third_party/cudnn_frontend'
2025-12-04T13:53:10.1032961Z Entering 'third_party/cutlass'
2025-12-04T13:53:10.1074915Z Entering 'third_party/fbgemm'
2025-12-04T13:53:10.1099884Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T13:53:10.1135152Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T13:53:10.1176421Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T13:53:10.1204150Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T13:53:10.1231924Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T13:53:10.1261115Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T13:53:10.1294412Z Entering 'third_party/fbgemm/external/json'
2025-12-04T13:53:10.1323874Z Entering 'third_party/flash-attention'
2025-12-04T13:53:10.1354132Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T13:53:10.1382644Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T13:53:10.1422901Z Entering 'third_party/flatbuffers'
2025-12-04T13:53:10.1452255Z Entering 'third_party/fmt'
2025-12-04T13:53:10.1475980Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T13:53:10.1502488Z Entering 'third_party/gloo'
2025-12-04T13:53:10.1532651Z Entering 'third_party/googletest'
2025-12-04T13:53:10.1563366Z Entering 'third_party/ideep'
2025-12-04T13:53:10.1587374Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T13:53:10.1623617Z Entering 'third_party/ittapi'
2025-12-04T13:53:10.1653356Z Entering 'third_party/kineto'
2025-12-04T13:53:10.1682074Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T13:53:10.1709419Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T13:53:10.1732138Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T13:53:10.1754684Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T13:53:10.1787595Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T13:53:10.1816118Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T13:53:10.1852695Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T13:53:10.1884433Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T13:53:10.1910631Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T13:53:10.1941963Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T13:53:10.1967483Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T13:53:10.1990695Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T13:53:10.2014344Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T13:53:10.2048487Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T13:53:10.2078917Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T13:53:10.2110693Z Entering 'third_party/kleidiai'
2025-12-04T13:53:10.2144686Z Entering 'third_party/mimalloc'
2025-12-04T13:53:10.2170121Z Entering 'third_party/nlohmann'
2025-12-04T13:53:10.2199983Z Entering 'third_party/onnx'
2025-12-04T13:53:10.2235016Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T13:53:10.2268471Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T13:53:10.2302199Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T13:53:10.2329852Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T13:53:10.2361943Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T13:53:10.2386988Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T13:53:10.2420915Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T13:53:10.2458072Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T13:53:10.2487355Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T13:53:10.2513323Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T13:53:10.2548027Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T13:53:10.2579466Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T13:53:10.2622785Z Entering 'third_party/pocketfft'
2025-12-04T13:53:10.2651879Z Entering 'third_party/protobuf'
2025-12-04T13:53:10.2679937Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T13:53:10.2707860Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T13:53:10.2738295Z Entering 'third_party/psimd'
2025-12-04T13:53:10.2765644Z Entering 'third_party/pthreadpool'
2025-12-04T13:53:10.2795742Z Entering 'third_party/pybind11'
2025-12-04T13:53:10.2819395Z Entering 'third_party/python-peachpy'
2025-12-04T13:53:10.2851646Z Entering 'third_party/sleef'
2025-12-04T13:53:10.2882103Z Entering 'third_party/tensorpipe'
2025-12-04T13:53:10.2909080Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T13:53:10.2945834Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T13:53:10.2977203Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T13:53:10.3005770Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T13:53:10.3034270Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T13:53:10.3080370Z [command]/usr/bin/git config --local --name-only --get-regexp http\.https\:\/\/github\.com\/\.extraheader
2025-12-04T13:53:10.3100780Z [command]/usr/bin/git submodule foreach --recursive sh -c "git config --local --name-only --get-regexp 'http\.https\:\/\/github\.com\/\.extraheader' && git config --local --unset-all 'http.https://github.com/.extraheader' || :"
2025-12-04T13:53:10.3288099Z Entering 'android/libs/fbjni'
2025-12-04T13:53:10.3314510Z Entering 'third_party/FP16'
2025-12-04T13:53:10.3348366Z Entering 'third_party/FXdiv'
2025-12-04T13:53:10.3380581Z Entering 'third_party/NNPACK'
2025-12-04T13:53:10.3406691Z Entering 'third_party/NVTX'
2025-12-04T13:53:10.3431674Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T13:53:10.3459738Z Entering 'third_party/XNNPACK'
2025-12-04T13:53:10.3498329Z Entering 'third_party/aiter'
2025-12-04T13:53:10.3522393Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T13:53:10.3551444Z Entering 'third_party/benchmark'
2025-12-04T13:53:10.3581262Z Entering 'third_party/composable_kernel'
2025-12-04T13:53:10.3615712Z Entering 'third_party/cpp-httplib'
2025-12-04T13:53:10.3648519Z Entering 'third_party/cpuinfo'
2025-12-04T13:53:10.3672437Z Entering 'third_party/cudnn_frontend'
2025-12-04T13:53:10.3698253Z Entering 'third_party/cutlass'
2025-12-04T13:53:10.3730983Z Entering 'third_party/fbgemm'
2025-12-04T13:53:10.3760091Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T13:53:10.3790612Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T13:53:10.3823117Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T13:53:10.3851533Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T13:53:10.3879425Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T13:53:10.3907644Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T13:53:10.3941395Z Entering 'third_party/fbgemm/external/json'
2025-12-04T13:53:10.3968361Z Entering 'third_party/flash-attention'
2025-12-04T13:53:10.3992771Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T13:53:10.4029017Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T13:53:10.4066245Z Entering 'third_party/flatbuffers'
2025-12-04T13:53:10.4092324Z Entering 'third_party/fmt'
2025-12-04T13:53:10.4115168Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T13:53:10.4141690Z Entering 'third_party/gloo'
2025-12-04T13:53:10.4166882Z Entering 'third_party/googletest'
2025-12-04T13:53:10.4191901Z Entering 'third_party/ideep'
2025-12-04T13:53:10.4217233Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T13:53:10.4250580Z Entering 'third_party/ittapi'
2025-12-04T13:53:10.4274603Z Entering 'third_party/kineto'
2025-12-04T13:53:10.4298524Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T13:53:10.4332139Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T13:53:10.4362566Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T13:53:10.4386851Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T13:53:10.4418534Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T13:53:10.4449403Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T13:53:10.4476079Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T13:53:10.4499462Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T13:53:10.4522220Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T13:53:10.4548776Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T13:53:10.4572276Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T13:53:10.4597654Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T13:53:10.4621425Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T13:53:10.4647646Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T13:53:10.4675434Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T13:53:10.4703108Z Entering 'third_party/kleidiai'
2025-12-04T13:53:10.4737492Z Entering 'third_party/mimalloc'
2025-12-04T13:53:10.4774959Z Entering 'third_party/nlohmann'
2025-12-04T13:53:10.4810199Z Entering 'third_party/onnx'
2025-12-04T13:53:10.4847197Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T13:53:10.4880188Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T13:53:10.4909178Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T13:53:10.4939912Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T13:53:10.4968140Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T13:53:10.4992344Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T13:53:10.5014649Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T13:53:10.5037665Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T13:53:10.5065307Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T13:53:10.5093801Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T13:53:10.5122553Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T13:53:10.5149015Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T13:53:10.5183781Z Entering 'third_party/pocketfft'
2025-12-04T13:53:10.5212162Z Entering 'third_party/protobuf'
2025-12-04T13:53:10.5240087Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T13:53:10.5269776Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T13:53:10.5302215Z Entering 'third_party/psimd'
2025-12-04T13:53:10.5327037Z Entering 'third_party/pthreadpool'
2025-12-04T13:53:10.5351917Z Entering 'third_party/pybind11'
2025-12-04T13:53:10.5373754Z Entering 'third_party/python-peachpy'
2025-12-04T13:53:10.5398515Z Entering 'third_party/sleef'
2025-12-04T13:53:10.5423723Z Entering 'third_party/tensorpipe'
2025-12-04T13:53:10.5450639Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T13:53:10.5470694Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T13:53:10.5500492Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T13:53:10.5532970Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T13:53:10.5557642Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T13:53:10.5604009Z [command]/usr/bin/git config --local --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.5627370Z [command]/usr/bin/git submodule foreach --recursive git config --local --show-origin --name-only --get-regexp remote.origin.url
2025-12-04T13:53:10.5796698Z Entering 'android/libs/fbjni'
2025-12-04T13:53:10.5817913Z file:/home/runner/_work/pytorch/pytorch/.git/modules/android/libs/fbjni/config	remote.origin.url
2025-12-04T13:53:10.5830382Z Entering 'third_party/FP16'
2025-12-04T13:53:10.5845350Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FP16/config	remote.origin.url
2025-12-04T13:53:10.5854529Z Entering 'third_party/FXdiv'
2025-12-04T13:53:10.5870251Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FXdiv/config	remote.origin.url
2025-12-04T13:53:10.5883772Z Entering 'third_party/NNPACK'
2025-12-04T13:53:10.5894972Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK/config	remote.origin.url
2025-12-04T13:53:10.5904920Z Entering 'third_party/NVTX'
2025-12-04T13:53:10.5914885Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NVTX/config	remote.origin.url
2025-12-04T13:53:10.5929615Z Entering 'third_party/VulkanMemoryAllocator'
2025-12-04T13:53:10.5941909Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/VulkanMemoryAllocator/config	remote.origin.url
2025-12-04T13:53:10.5951659Z Entering 'third_party/XNNPACK'
2025-12-04T13:53:10.5963527Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/XNNPACK/config	remote.origin.url
2025-12-04T13:53:10.5977444Z Entering 'third_party/aiter'
2025-12-04T13:53:10.5988102Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/config	remote.origin.url
2025-12-04T13:53:10.5997867Z Entering 'third_party/aiter/3rdparty/composable_kernel'
2025-12-04T13:53:10.6016158Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/modules/3rdparty/composable_kernel/config	remote.origin.url
2025-12-04T13:53:10.6030462Z Entering 'third_party/benchmark'
2025-12-04T13:53:10.6040483Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T13:53:10.6055133Z Entering 'third_party/composable_kernel'
2025-12-04T13:53:10.6065636Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/composable_kernel/config	remote.origin.url
2025-12-04T13:53:10.6080056Z Entering 'third_party/cpp-httplib'
2025-12-04T13:53:10.6091382Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpp-httplib/config	remote.origin.url
2025-12-04T13:53:10.6121357Z Entering 'third_party/cpuinfo'
2025-12-04T13:53:10.6121620Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpuinfo/config	remote.origin.url
2025-12-04T13:53:10.6130842Z Entering 'third_party/cudnn_frontend'
2025-12-04T13:53:10.6141647Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cudnn_frontend/config	remote.origin.url
2025-12-04T13:53:10.6149967Z Entering 'third_party/cutlass'
2025-12-04T13:53:10.6160328Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/cutlass/config	remote.origin.url
2025-12-04T13:53:10.6180250Z Entering 'third_party/fbgemm'
2025-12-04T13:53:10.6190021Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/config	remote.origin.url
2025-12-04T13:53:10.6206925Z Entering 'third_party/fbgemm/external/asmjit'
2025-12-04T13:53:10.6220497Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/asmjit/config	remote.origin.url
2025-12-04T13:53:10.6229372Z Entering 'third_party/fbgemm/external/composable_kernel'
2025-12-04T13:53:10.6239757Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/composable_kernel/config	remote.origin.url
2025-12-04T13:53:10.6252743Z Entering 'third_party/fbgemm/external/cpuinfo'
2025-12-04T13:53:10.6264694Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cpuinfo/config	remote.origin.url
2025-12-04T13:53:10.6275401Z Entering 'third_party/fbgemm/external/cutlass'
2025-12-04T13:53:10.6286073Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cutlass/config	remote.origin.url
2025-12-04T13:53:10.6303348Z Entering 'third_party/fbgemm/external/googletest'
2025-12-04T13:53:10.6315600Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/googletest/config	remote.origin.url
2025-12-04T13:53:10.6329142Z Entering 'third_party/fbgemm/external/hipify_torch'
2025-12-04T13:53:10.6345338Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/hipify_torch/config	remote.origin.url
2025-12-04T13:53:10.6354240Z Entering 'third_party/fbgemm/external/json'
2025-12-04T13:53:10.6370239Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/json/config	remote.origin.url
2025-12-04T13:53:10.6383197Z Entering 'third_party/flash-attention'
2025-12-04T13:53:10.6394124Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/config	remote.origin.url
2025-12-04T13:53:10.6402794Z Entering 'third_party/flash-attention/csrc/composable_kernel'
2025-12-04T13:53:10.6417770Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/composable_kernel/config	remote.origin.url
2025-12-04T13:53:10.6429925Z Entering 'third_party/flash-attention/csrc/cutlass'
2025-12-04T13:53:10.6446203Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/cutlass/config	remote.origin.url
2025-12-04T13:53:10.6464734Z Entering 'third_party/flatbuffers'
2025-12-04T13:53:10.6474795Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/flatbuffers/config	remote.origin.url
2025-12-04T13:53:10.6486088Z Entering 'third_party/fmt'
2025-12-04T13:53:10.6496046Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/fmt/config	remote.origin.url
2025-12-04T13:53:10.6505439Z Entering 'third_party/gemmlowp/gemmlowp'
2025-12-04T13:53:10.6515969Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/gemmlowp/gemmlowp/config	remote.origin.url
2025-12-04T13:53:10.6526859Z Entering 'third_party/gloo'
2025-12-04T13:53:10.6541953Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/gloo/config	remote.origin.url
2025-12-04T13:53:10.6551518Z Entering 'third_party/googletest'
2025-12-04T13:53:10.6561585Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/googletest/config	remote.origin.url
2025-12-04T13:53:10.6572903Z Entering 'third_party/ideep'
2025-12-04T13:53:10.6584291Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/config	remote.origin.url
2025-12-04T13:53:10.6593860Z Entering 'third_party/ideep/mkl-dnn'
2025-12-04T13:53:10.6603335Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/modules/mkl-dnn/config	remote.origin.url
2025-12-04T13:53:10.6617848Z Entering 'third_party/ittapi'
2025-12-04T13:53:10.6629608Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/ittapi/config	remote.origin.url
2025-12-04T13:53:10.6638568Z Entering 'third_party/kineto'
2025-12-04T13:53:10.6657673Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/config	remote.origin.url
2025-12-04T13:53:10.6670805Z Entering 'third_party/kineto/libkineto/third_party/dynolog'
2025-12-04T13:53:10.6683391Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/config	remote.origin.url
2025-12-04T13:53:10.6696244Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/DCGM'
2025-12-04T13:53:10.6710844Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/DCGM/config	remote.origin.url
2025-12-04T13:53:10.6720772Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/cpr'
2025-12-04T13:53:10.6732216Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/cpr/config	remote.origin.url
2025-12-04T13:53:10.6741003Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/fmt'
2025-12-04T13:53:10.6751823Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/fmt/config	remote.origin.url
2025-12-04T13:53:10.6760374Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags'
2025-12-04T13:53:10.6773047Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/config	remote.origin.url
2025-12-04T13:53:10.6783298Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/gflags/doc'
2025-12-04T13:53:10.6801020Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/modules/doc/config	remote.origin.url
2025-12-04T13:53:10.6811731Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/glog'
2025-12-04T13:53:10.6826702Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/glog/config	remote.origin.url
2025-12-04T13:53:10.6839142Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/googletest'
2025-12-04T13:53:10.6849976Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/googletest/config	remote.origin.url
2025-12-04T13:53:10.6859132Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/json'
2025-12-04T13:53:10.6870068Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/json/config	remote.origin.url
2025-12-04T13:53:10.6879399Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/pfs'
2025-12-04T13:53:10.6890318Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/pfs/config	remote.origin.url
2025-12-04T13:53:10.6898592Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp'
2025-12-04T13:53:10.6911466Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/config	remote.origin.url
2025-12-04T13:53:10.6920333Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T13:53:10.6930957Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/civetweb/config	remote.origin.url
2025-12-04T13:53:10.6944002Z Entering 'third_party/kineto/libkineto/third_party/dynolog/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T13:53:10.6955332Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/googletest/config	remote.origin.url
2025-12-04T13:53:10.6968029Z Entering 'third_party/kineto/libkineto/third_party/fmt'
2025-12-04T13:53:10.6977289Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/fmt/config	remote.origin.url
2025-12-04T13:53:10.6986343Z Entering 'third_party/kineto/libkineto/third_party/googletest'
2025-12-04T13:53:10.6995626Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/googletest/config	remote.origin.url
2025-12-04T13:53:10.7006365Z Entering 'third_party/kleidiai'
2025-12-04T13:53:10.7017178Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/kleidiai/config	remote.origin.url
2025-12-04T13:53:10.7026104Z Entering 'third_party/mimalloc'
2025-12-04T13:53:10.7036337Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/mimalloc/config	remote.origin.url
2025-12-04T13:53:10.7050888Z Entering 'third_party/nlohmann'
2025-12-04T13:53:10.7060393Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/nlohmann/config	remote.origin.url
2025-12-04T13:53:10.7069575Z Entering 'third_party/onnx'
2025-12-04T13:53:10.7079915Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/config	remote.origin.url
2025-12-04T13:53:10.7100383Z Entering 'third_party/onnx/third_party/pybind11'
2025-12-04T13:53:10.7117367Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T13:53:10.7131737Z Entering 'third_party/opentelemetry-cpp'
2025-12-04T13:53:10.7143127Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/config	remote.origin.url
2025-12-04T13:53:10.7152827Z Entering 'third_party/opentelemetry-cpp/third_party/benchmark'
2025-12-04T13:53:10.7162695Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T13:53:10.7171473Z Entering 'third_party/opentelemetry-cpp/third_party/googletest'
2025-12-04T13:53:10.7181837Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/googletest/config	remote.origin.url
2025-12-04T13:53:10.7190214Z Entering 'third_party/opentelemetry-cpp/third_party/ms-gsl'
2025-12-04T13:53:10.7205231Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/ms-gsl/config	remote.origin.url
2025-12-04T13:53:10.7213879Z Entering 'third_party/opentelemetry-cpp/third_party/nlohmann-json'
2025-12-04T13:53:10.7235938Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/nlohmann-json/config	remote.origin.url
2025-12-04T13:53:10.7246905Z Entering 'third_party/opentelemetry-cpp/third_party/opentelemetry-proto'
2025-12-04T13:53:10.7258676Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentelemetry-proto/config	remote.origin.url
2025-12-04T13:53:10.7269214Z Entering 'third_party/opentelemetry-cpp/third_party/opentracing-cpp'
2025-12-04T13:53:10.7286608Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentracing-cpp/config	remote.origin.url
2025-12-04T13:53:10.7294135Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp'
2025-12-04T13:53:10.7313890Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/config	remote.origin.url
2025-12-04T13:53:10.7323368Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/civetweb'
2025-12-04T13:53:10.7335907Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/civetweb/config	remote.origin.url
2025-12-04T13:53:10.7348585Z Entering 'third_party/opentelemetry-cpp/third_party/prometheus-cpp/3rdparty/googletest'
2025-12-04T13:53:10.7360978Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/googletest/config	remote.origin.url
2025-12-04T13:53:10.7375698Z Entering 'third_party/opentelemetry-cpp/tools/vcpkg'
2025-12-04T13:53:10.7386577Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/tools/vcpkg/config	remote.origin.url
2025-12-04T13:53:10.7408083Z Entering 'third_party/pocketfft'
2025-12-04T13:53:10.7418652Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/pocketfft/config	remote.origin.url
2025-12-04T13:53:10.7426741Z Entering 'third_party/protobuf'
2025-12-04T13:53:10.7436784Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/config	remote.origin.url
2025-12-04T13:53:10.7446871Z Entering 'third_party/protobuf/third_party/benchmark'
2025-12-04T13:53:10.7465487Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/benchmark/config	remote.origin.url
2025-12-04T13:53:10.7477948Z Entering 'third_party/protobuf/third_party/googletest'
2025-12-04T13:53:10.7498180Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/googletest/config	remote.origin.url
2025-12-04T13:53:10.7510845Z Entering 'third_party/psimd'
2025-12-04T13:53:10.7524028Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/psimd/config	remote.origin.url
2025-12-04T13:53:10.7533136Z Entering 'third_party/pthreadpool'
2025-12-04T13:53:10.7544012Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/pthreadpool/config	remote.origin.url
2025-12-04T13:53:10.7553248Z Entering 'third_party/pybind11'
2025-12-04T13:53:10.7569162Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T13:53:10.7579950Z Entering 'third_party/python-peachpy'
2025-12-04T13:53:10.7590561Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/python-peachpy/config	remote.origin.url
2025-12-04T13:53:10.7599244Z Entering 'third_party/sleef'
2025-12-04T13:53:10.7608634Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/sleef/config	remote.origin.url
2025-12-04T13:53:10.7620487Z Entering 'third_party/tensorpipe'
2025-12-04T13:53:10.7630798Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/config	remote.origin.url
2025-12-04T13:53:10.7639664Z Entering 'third_party/tensorpipe/third_party/googletest'
2025-12-04T13:53:10.7648689Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/googletest/config	remote.origin.url
2025-12-04T13:53:10.7657611Z Entering 'third_party/tensorpipe/third_party/libnop'
2025-12-04T13:53:10.7667877Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libnop/config	remote.origin.url
2025-12-04T13:53:10.7680087Z Entering 'third_party/tensorpipe/third_party/libuv'
2025-12-04T13:53:10.7690038Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libuv/config	remote.origin.url
2025-12-04T13:53:10.7699984Z Entering 'third_party/tensorpipe/third_party/pybind11'
2025-12-04T13:53:10.7709829Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/config	remote.origin.url
2025-12-04T13:53:10.7719225Z Entering 'third_party/tensorpipe/third_party/pybind11/tools/clang'
2025-12-04T13:53:10.7729559Z file:/home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/modules/tools/clang/config	remote.origin.url
2025-12-04T13:53:10.7756458Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/android/libs/fbjni/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.7777148Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FP16/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.7794820Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/FXdiv/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.7811414Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.7833495Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NVTX/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.7852089Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/VulkanMemoryAllocator/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.7868503Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/XNNPACK/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.7885338Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.7902402Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/aiter/modules/3rdparty/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.7922915Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/benchmark/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.7942166Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.7958425Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpp-httplib/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.7978901Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cpuinfo/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.7995035Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cudnn_frontend/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8009749Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/cutlass/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8029161Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8046135Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/asmjit/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8065621Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8082803Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cpuinfo/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8098181Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/cutlass/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8118774Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8139223Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/hipify_torch/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8157614Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fbgemm/modules/external/json/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8175040Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8190594Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/composable_kernel/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8206907Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flash-attention/modules/csrc/cutlass/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8226575Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/flatbuffers/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8242808Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/fmt/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8260513Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/gemmlowp/gemmlowp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8285595Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/gloo/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8301361Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8322855Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8340833Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/ideep/modules/mkl-dnn/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8362354Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/ittapi/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8384363Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8402770Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8421440Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/DCGM/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8444224Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/cpr/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8464249Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/fmt/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8484225Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8506781Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/gflags/modules/doc/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8524042Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/glog/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8543616Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8568908Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/json/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8586677Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/pfs/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8604956Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8624211Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/civetweb/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8640559Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/dynolog/modules/third_party/prometheus-cpp/modules/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8657857Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/fmt/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8676333Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kineto/modules/libkineto/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8692394Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/kleidiai/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8708236Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/mimalloc/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8725759Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/nlohmann/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8742505Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8759547Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/onnx/modules/third_party/pybind11/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8776011Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8796234Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/benchmark/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8817713Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8835360Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/ms-gsl/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8852159Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/nlohmann-json/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8870574Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentelemetry-proto/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8887903Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/opentracing-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8905154Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8922034Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/civetweb/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8946948Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/third_party/prometheus-cpp/modules/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8966771Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/opentelemetry-cpp/modules/tools/vcpkg/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.8985177Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/pocketfft/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9002471Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9022375Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/benchmark/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9039393Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/protobuf/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9060825Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/psimd/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9080353Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/NNPACK_deps/pthreadpool/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9103326Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/pybind11/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9120387Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/python-peachpy/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9137827Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/sleef/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9156242Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9173130Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/googletest/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9190383Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libnop/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9207796Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/libuv/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9225405Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9245636Z [command]/usr/bin/git config --file /home/runner/_work/pytorch/pytorch/.git/modules/third_party/tensorpipe/modules/third_party/pybind11/modules/tools/clang/config --name-only --get-regexp ^includeIf\.gitdir:
2025-12-04T13:53:10.9364503Z Cleaning up orphan processes
2025-12-04T13:53:10.9437530Z Terminate orphan process: pid (17343) (docker)